Share

Regarding WAICO and the AI Governance Summit: On June 17, Xinhua reported that China will host the 2026 World Artificial Intelligence Conference and High-Level Meeting on Global AI Governance in Shanghai this July. A State Council Information Office press conference the same day confirmed Beijing is accelerating WAICO’s formal establishment, with Shanghai as its planned headquarters. WAICO now has a calendar date. Western governments do not have a comparable institutional vehicle to bring to the same table.

Commerce ministry issues 17 AI-consumption integration measures. China’s Ministry of Commerce announced 17 specific measures on June 18 to embed AI across household and business consumption sectors nationwide, per Reuters. The package covers both AI products and services and is designed to push AI adoption from the model layer down into everyday commerce. This is the domestic counterpart to the international governance push Beijing announced the day before.

Soufan Center: AI model superiority over PRC is a live national security priority. The Soufan Center’s June 17 intelligence brief analyzed the U.S. government’s decision to block Anthropic’s Fable 5 model for foreign nationals over national security concerns. The brief documents Anthropic’s November 2025 disclosure of a Chinese state-sponsored group that used Claude to execute an AI-orchestrated cyberattack, with the model completing 80 to 90 percent of the operation autonomously against technology companies and government agencies. The brief notes that advanced AI models now carry documented biological research competence applicable to weapons development and demonstrated attack-surface auditing capability applicable to offensive cyber operations.

War on the Rocks: Chinese open-weight models now lead global downloads. A War on the Rocks analysis published this week found that Chinese open-weight models accounted for 17.1 percent of global AI model downloads in recent measurement periods, surpassing the U.S. share of 15.9 percent for the first time. The piece argues that China’s AI progress is driven as much by market competition, specifically the domestic “involution” dynamic known as neijuan, as by state industrial policy. The data point that should alarm policymakers: roughly 80 percent of U.S. startups currently use Chinese base models, not by mandate, but because Chinese models are cheaper and more accessible. That market penetration is operating below the reach of hardware export controls.

DeepSeek’s governance structure formalizes the state alignment risk. China’s National Intelligence Law (2017, Article 7), Cybersecurity Law (2017), and Data Security Law (2021) collectively create legal obligations for Chinese companies to cooperate with government data and security requests. The Carnegie Endowment concluded in January 2025 that Chinese companies are subject to legal regimes that can compel cooperation with Chinese defense and intelligence services. The June 16 funding structure resolves any ambiguity about DeepSeek’s position within that framework. The state AI fund holds the only vote. Every enterprise or government agency building on DeepSeek’s open-weight releases should treat that governance fact as a known variable in their risk calculus.

Strategic Assessment

The DeepSeek funding structure closed Tuesday with an answer to a question that has structured a year of U.S. China AI policy debate. Beijing holds the only vote. Enterprise developers building on V4 weights are working with technology whose parent company is formally controlled by the Chinese state and the CCP. At the cost frontier, V4 is 29 times cheaper per output token than comparable closed-source alternatives. For workloads that fall within V4’s capability band, the economics are compelling. The governance risk is now explicit.

WAICO’s July timeline is most likely to be undercounted by policymakers. Beijing has spent 12 months building the institutional architecture for a global AI governance body. It arrives at Shanghai in July with a proposal, a host city, and backing from Global South nations who have been deliberately recruited through an equity and development framing. What the West has is a set of uncoordinated national processes. That asymmetry in preparation will show in July, and what gets established at Shanghai will carry normative weight regardless of whether the U.S. formally participates.

The open-weight download crossing is a lagging indicator of a trend that has been building since DeepSeek V3 hit Hugging Face. Eighty percent of U.S. startups use Chinese base models because price competition, which involution has driven to a floor that U.S. labs cannot match without compromising their business model, has already determined the choice at the application layer. Export controls address the hardware supply chain. No current U.S. policy instrument addresses the adoption supply chain already in place.

Share

Beijing is running a two-track strategy: flood global markets with open-source AI capability, then use the venue they established, the World AI Cooperation Organization (WAICO), to write the international rules that govern it. Congress should pay attention. So should you.

Share

What Is This WAICO Thing?

The World AI Cooperation Organization, proposed by China in mid-2025, has matured from a diplomatic talking point into an active institutional project which is being weaponized against the U.S. Its architecture, including a technology-sharing platform, equity adjustment mechanism, and algorithmic compensation fund is designed to attract Global South nations by framing AI governance as a development equity issue. Carnegie Endowment’s analysis calls this a deliberate pivot: Beijing wants to be the standards-setting body, not just the technology exporter. (Background on WAICO’s proposed structure is available from CGTN and the North American Academy of Sciences).

Some recent PRC AI news:

Alibaba pivots to physical AI, announces robotics-native models. Alibaba this week moved beyond large language models to release AI systems explicitly designed for robotic and physical-world applications a direct play for the industrial and manufacturing sectors where China already holds significant supply chain advantages. The move accelerates Beijing’s civil-military fusion agenda in robotics and autonomous systems.

Beijing compute voucher program in active enrollment. Local governments across China’s major tech hubs are enrolling AI developers in a compute subsidy program designed to attract domestic talent and accelerate model development. Lawfare’s analysis notes the contrast with Washington: while U.S. export controls targeted chips and hardware, Chinese municipalities competed to distribute rewards. The incentive architecture operates below the threshold export controls can reach.

USCC “Two Loops” report identifies China’s open AI strategy as industrial dominance play. The U.S.-China Economic and Security Review Commission published a detailed analysis of how Beijing’s open-source model releases function as a tool of industrial strategy not open-science altruism. Chinese labs release powerful open-weight models that drive global developer adoption at the application layer, entrenching Chinese AI infrastructure in enterprises and governments worldwide.

China and U.S. both decline military AI joint declaration. At a February 2026 summit, approximately one-third of attending nations signed a joint declaration governing AI use in warfare. Both China and the United States opted out for different reasons, with different implications. China’s refusal signals it will not accept externally imposed constraints on PLA AI integration.

PLA “intelligentization” drive expands across services. Defense News reported in April that AI is now expanding across PLA branches under a doctrine explicitly framing “intelligentization” as preparation for potential South China Sea and Taiwan Strait conflicts. Chinese PLA-affiliated tech firms, including CETC, are integrating surveillance data and AI for multi-platform missile strike coordination.

Carnegie Endowment: Beijing’s AI diplomacy has pivoted from infrastructure exports to governance architecture. Carnegie’s May 2026 analysis found that Beijing has shifted from a first-generation strategy of exporting Chinese hardware and standards toward a more comprehensive second-generation effort to reshape the norms and institutions that govern AI globally. This is a significant strategic escalation.

Compute subsidy programs as export control countermeasure. Beijing’s local government compute voucher programs represent a structural response to U.S. chip restrictions not by acquiring restricted hardware, but by accelerating domestic AI development through financial incentives. Lawfare and CFR both note that this incentive architecture operates outside the reach of hardware export controls.

Military AI without guardrails. Both China and the U.S. declined the February 2026 joint military AI declaration but from asymmetric positions. Washington is debating internal governance frameworks; Beijing has issued no public AI safety constraints on PLA applications and is actively integrating AI into multi-domain strike coordination systems.

Export controls debate continues. CFR’s January 2026 analysis of the Trump administration’s revised AI chip export policy notes that hardware controls alone are strategically insufficient when Chinese labs are demonstrating frontier performance on domestically available compute. Chatham House reached the same conclusion in April: chip restrictions cannot prevent China from further developing advanced AI.

I built a tool that can help. I could use your help testing it out so I can ensure it is ready for prime time.

Here is what I am trying to do:

- I want any small business to be able to use this tool to learn about the security posture of tech associated with their domain, and want it to be free for any small business to use.

- When a domain name is entered in the tool, it conducts scans and configuration checks on the domain and queries several databases. The system uses tools most security practitioners would know, but few small business owners would ever think about using.

- Rather than just dumping results on them, the scanning results are summarized and converted into plain language. The goal is to produce something that any small business owner would be able to understand.

My request:

If you know cybersecurity, please test the tool out and let me know what you think of the results. It has been tested by a small group of friends but I want to widen the circle of people who can kick the tires and give me feedback.

If you are not a security pro, would appreciate you testing and giving me feedback on how results are expressed. Test any domain and let me know what you think.

See https://securityscans.ai

Enter your domain. Run a scan or two. Check the results.

I’m looking for honest feedback. Try it on your domain or forward it to a small business owner you know. Is the report useful? Are the recommendations clear? What would make it better?

Comments and DMs welcome. The more direct, the better.

Subscribe now

In the Exponential and Adolescence posts he calls on Congress to mandate independent safety testing for frontier AI models above a compute threshold, and advocates for government authority to block releases that fail those audits.

He wants government control with teeth. I believe he also wants to destroy his enemies.

The AI Exponential post itself is a great read. Anything Dario writes is worthy of review for sure. But we are all thinkers and deserve to form our own opinions on topics of such critical importance.

Share

His argument: AI policy is running way behind where the tech is, and that gap is dangerous enough to require a more serious response. (welcome to my world! I have witnessed this struggle for decades now!).

The piece opens with a Lord of the Rings analogy. AI is moving at hobbit speed while governments move like Treebeard. Not sure we needed the analogy, could have just stated the fact, the government is not designed to legislate at speed.

There is plenty of evidence that the government will never catch up to what society needs from them when it comes to tech regulation. They have never caught up to what is needed in cybersecurity, or the age of big data, or smartphones. This has been clear on the topic of AI from the beginning, and now Claude Mythos and its ability to discover vulnerabilities underscored that again.

The five areas he addresses:

Regulation. He wants mandatory pre-release third-party testing for frontier models in four specific risk categories: cybersecurity, bioweapons, loss of control, and automated R&D that could accelerate the others. The government would have the power to block or delay deployment on safety grounds. He uses the FAA as his model, not the FDA.

Economics and labor. He’s direct that significant enduring job displacement is possible and might be intrinsic to a technology that replicates general cognition rather than just automating a narrow task. He calls for better measurement, pro-employment tax incentives, and long-term income support mechanisms (including UBI financed by taxes on AI companies and capital gains) if displacement turns out to be large and permanent. He separates the economic problem from the deeper question of meaning and purpose, noting policy can buy time but can’t answer the second one.

Accelerating AI’s benefits. For downstream fields like biomedicine, his concern runs the other direction. He understands that regulatory systems designed for a slower pace of innovation will jam up and block the benefits. He wants the FDA and EMA to start now developing standards for AI-based simulation methods so they can be adopted quickly when they work.

State power and civil liberties. He is concerned that AI could become the ultimate autocracy tool. He wants mandatory accountability mechanisms for autonomous weapons, a ban on fully autonomous weapons in domestic law enforcement, closure of the bulk data/data-broker surveillance loophole, and a principle that anyone subject to adverse government action should have access to AI at least as capable as what the government is using against them.

Geopolitics. He seems to believe that AI is closer to nuclear weapons than to trade policy. A nation with powerful AI facing one without it could be the equivalent of World War II Marines against medieval swordsmen. He wants a democratic coalition that shares chips and semiconductor equipment among members, coordinates on safety standards, commits to mutual defense, and explicitly rejects AI-powered repression as a membership condition. The goal is to make being outside the coalition progressively more costly.

This is evolved from his previous posts on The Adolescence of Technology. That saw him set up what I think are a few strawman arguments so he can knock them down. Like doomerism. He thinks it is bad now. In criticizing it he makes it seem like he is not one of those! But then brings up the doomer arguments to claim we must control this stuff with excessive government regulation. He also argues that government regulation must only be surgical since government action can destroy economic value and make AI worse for us all. After saying that he goes on to argue that we must have hard core government regulation. My head is spinning over that about face!

Here’s what the coverage mostly missed: Amodei published this essay one day after Anthropic released Claude Fable 5 (see this Fable review for some good and not so good with this model). This is Anthropic’s most capable public model. It is a gated version of the infamous Claude Mythos, the system Dario cites in the essay as evidence that cybersecurity risks have already materialized. That’s not a coincidence in timing. It’s a product launch and a regulatory brief delivered in the same 24-hour window.

Subscribe now

The core argument is this: Claude Mythos Preview demonstrated it could discover thousands of high-severity zero-days across every major OS and most major applications including every browser. That capability is real. The risks are real. Therefore, government should require mandatory third-party audits before any model of comparable power ships, with binding authority to block deployment if the auditors say no.

Among the many structural questions embedded in his argument that Amodei doesn’t answer: who runs the audits?

The Trump executive order that preceded this essay encouraged voluntary model-sharing with government auditors a month before public release. Amodei wants something stronger, mandatory testing with binding consequences. That’s a meaningful governance upgrade. But the audit infrastructure doesn’t exist yet. The only organizations currently capable of evaluating frontier models at this level of sophistication are... the frontier labs themselves. NIST’s AI evaluation center just went dark while the executive order gets implemented. The independent third-party ecosystem is nascent.

Which means the most likely near-term outcome of mandatory audit requirements isn’t an independent safety regime. It’s a certification framework that the labs with the most resources, Anthropic, OpenAI, Google DeepMind, Meta, help design, staff, and effectively control. Smaller competitors don’t have the safety teams to navigate that process. This builds the biggest deepest moat ever dug!

None of this means the underlying case for mandatory testing is wrong. Cybersecurity risks from frontier models are real and accelerating. The patch-one-CVE-at-a-time model of federal vulnerability management was already straining before AI worms entered the picture. The question of who controls the epistemic infrastructure of AI governance, like who decides what “safe enough” means, is one of the defining policy questions of the next decade.

But the people raising the alarm most loudly right now are the same people whose business models benefit most from the answer.

Questions worth sitting with:

• If Anthropic designed the audit framework, funded the auditors, and its own models set the capability benchmarks, is that a safety regime or a standards cartel?

• At what point does “we’re calling for regulation” become “we’re calling for the regulation we wrote”?

• Claude Mythos can find thousands of zero-days. Who has access to that output right now, and under what terms?

• If the federal government can block a model release, who in government makes that call, and what’s their threat model? Are they qualified to make calls like that?

• Trump’s executive order went in the direction of voluntary disclosure. Amodei wants mandatory blocking authority. What’s the actual path through Congress for that, given the current AI-competitiveness-vs-China framing?

• Anthropic released its most capable model the day before calling for mandatory testing of most-capable models. What does that sequencing tell you about the actual stakes?

Subscribe now

Share

I’ve been watching AI development and the national security implications of it closely for years. I’ve sat through the post-2023 pause letters, tracked the doomer arguments, and watched that movement fail to stop anything. The argument now is more sophisticated, and it’s coming from a more credible institution. But my read hasn’t changed. I still believe Accelerating AI is a net good for humanity.

This new report and call for a pause is not Anthropic acting in the best interests of humanity. This is Anthropic acting from a position of strength, proposing governance architecture that would cement their lead while their IPO paperwork sits at the SEC.

Let me explain what I mean, and then I want to spend some time on the political philosophy here, because the arguments Anthropic hasn’t made are more interesting than the ones they have.

The Timing Problem: Seems to look like this is all about the money

Anthropic filed confidentially for an IPO on June 2, 2026, two days before this report came out. The filing targets an October 2026 listing and sits on a $965 billion valuation. Run-rate revenue is approaching $47 billion annually, up from $9 billion at the end of 2025. Enterprise customers spending over $1 million a year now number more than 1,000, double the figure from two months prior. Eight of the Fortune 10 are paying customers.

Does that sound like the financial profile of a company worried it won’t survive? That’s a company locking in its position before the public markets open.

The pause proposal, if somehow it worked, would benefit Anthropic in at least two ways. First, it freezes the current capability hierarchy. Anthropic is at or near the frontier. Stopping development at this moment means stopping development with Anthropic’s models in the enterprise workflows of the world’s largest companies. And it hurts innovative startups that could one day challenge Anthropic. Second, it positions Anthropic as if they are the responsible adult in the room, which matters for enterprise sales, regulatory relationships, and the narrative heading into an IPO. Every conversation about “responsible AI” is a conversation Anthropic wins on brand, because they’ve spent years investing in that identity.

I want to be clear: I don’t think Dario Amodei and his team are cynical actors. I even believe they believe the safety concerns they’re articulating (I don’t, not really). But believing something sincerely doesn’t mean the structural incentives aren’t also real. Both can be true. The incentives are real.

The Hobbes problem

Anthropic’s core ask fails the Hobbesian test immediately. Thomas Hobbes built his entire political philosophy on one observation: without a sovereign with real enforcement power, agreements among self-interested actors collapse. The Leviathan isn’t metaphorical. It’s the structural requirement for cooperative restraint to hold.

Anthropic is asking for a global Leviathan to govern AI development while simultaneously acknowledging that no such entity exists. The nuclear analogy they invoke actually demonstrates the problem. Arms control worked slowly, imperfectly, and only between two superpowers who had both already experienced the catastrophic demonstration of what they were containing. Mutual assured destruction created the Hobbesian brake. We haven’t had that event in AI. The shared trauma that makes states serious about coordination hasn’t happened yet.

There’s no international body with the authority, the verification capability, or the enforcement mechanism to make a pause real. The United Nations can’t do it. No bilateral treaty framework covers it. The PRC’s AI development has its own political imperatives that no external governance proposal addresses.

Without the Leviathan, the pause proposal is a call for voluntary restraint among competitive actors. Voluntary restraint among competitive actors is just unilateral disarmament for whoever takes it seriously.

Kant’s categorical imperative, broken by structure

Kant’s test asks whether a maxim can be universalized. “Stop building the most powerful AI systems” cannot be universalized, not only because it’s a wrong principle, but because the competitive structure makes universalization impossible without the Hobbesian sovereign. Any actor who applies the maxim unilaterally ends up at catastrophic disadvantage relative to any actor who doesn’t.

Anthropic knows this. Their report says it plainly: if only one company stops, rivals race ahead. They’re not calling for individual voluntary restraint. They’re calling for an institutional solution to a structural coordination failure. That, in my opinion, is not the correct framing of the problem. But even if it were, this type of maxim cannot be universalized. Anthropic’s elegance can’t conjure a controlling authority or institution.

What we’re left with is a correctly diagnosed problem and no mechanism for the cure.

Plato’s missing guardian class

Plato’s deepest concern in The Republic wasn’t about the form of government per se. It was about who should hold dangerous knowledge and power. His answer: those who both understand it deeply and have demonstrated the character to subordinate their own interests to the common good. His nightmare was dangerous capability in the hands of people with technical mastery but no such formation.

Read what Anthropic is actually saying: they’ve built Mythos, they’ve seen what it can do, and they don’t fully trust anyone, including themselves, with the complete capability set. That’s not false modesty. That’s a genuine Platonic crisis. The knowledge exists. The guardian class doesn’t.

Project Glasswing, which I’ve written about before in the context of what Mythos found in open-source code, is Anthropic’s attempt to approximate the guardian function, carefully routing the model’s most dangerous capabilities only toward vetted defensive purposes. That’s a serious effort. It doesn’t scale to a governance framework for global AI development, but it’s an honest attempt to operate responsibly given real constraints.

What the strategists would say

Andy Grove called moments like this Strategic Inflection Points, 10X forces that change the competitive map irreversibly. His prescription at such moments wasn’t consensus-building. It was decisive movement through the point, because hesitation destroys more value than commitment. Anthropic has correctly identified the inflection. Their prescribed response is wrong. You don’t build the coalition while the inflection is active. You move through it, set the new terms, then govern from a position of demonstrated leadership.

Clayton Christensen would notice something structurally odd here. Incumbents normally resist disruptive change. Here, some of the most advanced actors are requesting governance constraints on the disruption they themselves are generating. That’s the innovator’s dilemma running in reverse. Christensen would ask: is this genuine safety concern or moat-building? The honest answer is both, and the distinction may matter less than the structure.

If Anthropic succeeds in helping construct the regulatory framework governing frontier AI, that framework will almost certainly be shaped in ways that favor companies with existing safety infrastructure, established government relationships, and the resources to navigate compliance requirements. That’s how incumbents shape regulation everywhere- it is regulatory capture. The observation matters because we should evaluate the proposal accordingly.

What I actually think should happen

I’m not arguing for zero governance. I want serious thinking about AI risks and smart mitigation of risks (all risks I have seen can be mitigated). And I love the power Anthropic’s capability gives me and my business and businesses across the nation. And I get that there should be governance.

What I don’t want is for the United States to slow down at this moment. Not unilaterally. Not while the PRC is running their own frontier development programs with no intention of pausing. Not while open-weight models from Chinese firms are spreading across global markets. Not while AI capability is the central variable in what my readers in national security and defense understand is the defining strategic competition of the next several decades.

I’ve said before, and I’ll keep saying it: I would like to see all PRC AI banned in U.S. markets. That position doesn’t get softer when Anthropic asks me to consider slowing down American development.

The doomers have been trying to pause AI research since at least 2023. I was glad when that effort failed. I’m glad when it fails now. The lesson from every inflection point in technology history is that the actors who move through it with skill and speed set the terms for what comes after. The actors who pause waiting for consensus get left out of that conversation entirely.

Anthropic has built extraordinary things. It gives me a bit of a cognitive dissonance to say this, but I love them. But that does not change my assessment of what their recent proposal is: a well-resourced, well-intentioned company trying to shape the governance environment in ways that serve their interests, dressed in language about humanity.

My view remains: Accelerate thoughtfully. Yes lets identify risks and mitigate them. But Don’t Pause!

If the strategic competition framing here interests you, the scenarios I’ve been working through in the military techno-thriller Dragon Falls get at why this moment matters more than most people outside the national security community currently understand. Would love it if you take a look.

Subscribe now

Share

I’ve watched this pattern long enough to know that the scores are telling us less and less about what matters.

Subscribe now

This past week I’ve been tracking the latest wave of AI industry data and one finding stopped me cold. A May 2026 study took 510 domain-specific questions, judged the answers with licensed professionals in relevant fields, and found that GPT-5.5, Claude Opus 4.7, and Gemini 3 Pro all scored in the 67-73% range. C-plus territory. And here’s what makes that genuinely important: adding more inference compute barely moved the needle. The models that command trillion-dollar valuations and run on infrastructure costing hundreds of billions in annual capital expenditure are, when professionals put them to the test on real work, scoring at roughly the level of a passing grade in a community college course.

That result should reset how everyone in our community thinks about AI capability assessment.

The benchmark problem

Benchmark scores were always a proxy. They measured something real -- the ability of a model to navigate structured question-and-answer formats, to handle math problems, to complete code snippets in isolation. For a few years that was useful. The distance between GPT-3 and GPT-4 on benchmarks corresponded to something meaningful in practice.

That correspondence is breaking down.

The reason is that benchmarks measure isolated competence. A model can ace a bar exam practice test and still fail a lawyer when the task requires synthesizing ambiguous case law, managing a client relationship, making a judgment call under time pressure, and knowing what questions to ask when the facts don’t fit the framework. Professional judgment isn’t a bigger version of question answering. It’s a different cognitive act.

What the May study found is that frontier models are hitting a wall specifically on professional judgment, prioritization, and high-stakes decisions in health and law. Those aren’t edge cases. Those are the use cases that enterprises are betting billions on.

It’s worth noting how we got here. MMLU -- Massive Multitask Language Understanding, introduced by Dan Hendrycks and colleagues in 2020 -- was a genuine advance when it appeared. Fifty-seven subjects, from STEM to law to history, tested in zero-shot and few-shot settings. It was hard. The largest GPT-3 models barely beat random chance on it. Then models got better, and MMLU got saturated. HLE (Humanity’s Last Exam) was created by the Center for AI Safety and Scale AI specifically because models were hitting 90%+ on MMLU and the benchmark had stopped being informative. HLE is harder -- 2,500 expert-crafted questions, many with no answer findable via web search. The top models now score around 45% on it. Within two years, they’ll saturate that too. Benchmark creation is a treadmill. The real world doesn’t work that way.

What the right evaluation looks like

I’ve been arguing for a while that we need to move to a more complete framework for assessing AI capability. The dimensions that matter now are architecture, harness, and diffusion.

Architecture is the model itself -- but evaluated on real task completion, not benchmark scores. Can it handle ambiguity? Does it know what it doesn’t know? How does it perform on the actual work of your domain, tested by the people who do that work? The Forbes study cited above is a step in the right direction. More evaluations like that, in more domains, run by practitioners rather than benchmark committees.

Harness is how the model is deployed. A capable model running in a bad harness produces bad outcomes. A less capable model running in a well-designed agentic workflow, with the right tools and memory and retrieval, can outperform the frontier model sitting in a chat window. The explosion of agentic frameworks -- the work being done in OpenClaw, in Microsoft’s Copilot Studio computer-use GA, in Anthropic’s self-hosted sandboxes -- is where the practical differentiation is happening. By some counts, 97% of executives report deploying AI agents in the past year. What varies enormously is the quality of the harness around those agents.

Diffusion is whether the capability is actually reaching the people and organizations that need it. This is where the numbers get sobering. U.S. Census Bureau data from December 2025 through May 2026 shows business AI usage hovering between 17-20%. Enterprise spending has nearly doubled -- companies are now spending 1.7% of revenue on AI, up from 0.8% last year -- but only 29% report significant ROI from generative AI. The headline that captures the moment: 79% of organizations say they’re struggling.

That gap between investment and returns is a diffusion problem. The companies that have figured out deployment are doing remarkable things -- Salesforce claims 20,000 engineers on AI coding tools at over 90% active usage. Most companies are nowhere near that. The capability is available. Getting it to work at scale, inside real organizational structures, with real governance requirements, is the hard part.

The PRC dimension that isn’t going away

I’d be incomplete if I didn’t note what this picture looks like from a strategic competition standpoint.

Chinese open-weight models -- DeepSeek V4, Kimi K2.6, Qwen3 -- now trail closed frontier models by roughly 6 points on composite intelligence indices while costing a fraction of the price. For cost-sensitive workloads, that gap is shrinking to the point where the procurement decision is no longer obvious. DeepSeek V4 matches or beats proprietary models on coding benchmarks. And it now runs natively on Huawei Ascend chips -- the first frontier-class model trained and deployed entirely on Chinese domestic semiconductor infrastructure. The 2026 AI Index from Stanford HAI confirmed what anyone watching closely already knew: the U.S.-China model gap has closed.

I’ve written before about the PRC’s approach to AI competition. My novel Dragon Falls is built around exactly this dynamic -- how Beijing weaponizes American-developed technology and turns it against American interests. That’s fiction grounded in what I’ve seen across my career in intelligence and at OODA. The open-weight model strategy is real-world confirmation of that playbook. DeepSeek V4 on Chinese chips, distributed globally under an open license, diffusing into global enterprise infrastructure -- that’s not a benchmark story. That’s a strategic access story.

I would like to see all PRC AI banned from U.S. markets. I’ve said that before. The economic case for using cheap Chinese models is clear in the short run. The strategic risks are not being priced in.

What the valuation picture is actually telling us

Anthropic just closed a Series H at approximately $965 billion, overtaking OpenAI to become the most valuable private AI startup in the world. The timing of a late 2026 IPO is on the table. At the same time, Microsoft scaled back internal Claude licenses when costs hit $500-2,000 per engineer per month at high adoption. Anthropic is reportedly in talks to rent Microsoft-designed chips because its own compute costs are unsustainable.

Those two facts together reveal the central tension of the AI industry right now. Valuations are priced for a future in which these models become deeply embedded in enterprise workflows and generate commensurate returns. The unit economics of getting there are brutal. The $500-2,000 monthly per power user isn’t sustainable at scale without dramatic reductions in model cost. The good news is that inference costs are dropping roughly 10x per year for equivalent capability. NVIDIA’s Vera Rubin platform, shipping H2 2026, claims up to 10x lower inference token costs. AMD’s MI400 series on TSMC 2nm is competing directly. The hardware war is now fought at the rack level.

If inference economics follow the trajectory they’ve been on, models that cost $30 per frontier query today cost $3 in a year. That changes which applications make sense. It also means today’s “too expensive” becomes tomorrow’s standard operating procedure.

What to watch

The question for the next 12 months isn’t which model scores highest on the next benchmark. But then what is it? I’m still looking for the right metric and have not found it. But believe we can frame what we need and at least scope the problem better. Here are three things to watch:

First, whether agentic workflows can actually close the ROI gap. The 29% of enterprises seeing significant returns from generative AI need to become 60% or 70%. The architecture-harness-diffusion framework is the right lens for tracking that. When you hear a company announce AI results, ask: what was the harness? What was the deployment model? What did professionals actually test?

Second, whether PRC open-weight models continue to close the capability gap at a fraction of Western cost. If DeepSeek V5 matches GPT-6 on professional task completion at 10% of the price, the procurement calculus changes dramatically -- and with it, the strategic access picture.

Third, whether any regulation actually bites. The EU just pushed high-risk AI obligations back 16 months, to December 2027. State-level requirements in Illinois, California, and New York are creating compliance overhead that advantages well-funded labs over smaller competitors. The federal picture in the U.S. remains fragmented. I covered the AI Action Plan when it came out -- the strategic logic was sound, even if implementation will be messy. Regulation is moving slower than capability. That’s both an opportunity and a risk.

The bottom line

Stop optimizing for benchmark scores. Start building evaluation frameworks that test what matters: professional task completion, real domain performance, quality of the harness, speed of diffusion into your organization. The models that score 70% on professional judgment today will score 80% in a year. The organizations that figure out deployment now will have compounding advantages that no future model upgrade can eliminate overnight.

The OODA Network members I talk to who are seeing real returns from AI have one thing in common: they stopped waiting for the models to get better and started building to integrate capabilities into their workflows. Playing is absolutely the best way to learn. Build something, break it, learn what works.

For deeper analysis on the evolving strategic AI competition, OODA Loop has more: oodaloop.com.

Share

Share

That was the spark that drove me to write Dragon Falls. I felt a need to tell a story that put readers inside the reality of what PRC covert operations look like when they’re running at full scale, because too few people understand the scope of what is happening. The novel is fiction. The threat is not.

What the FBI found on East Broadway

The federal indictment laid out the operation in plain terms. Lu and Chen opened the office space around 2022 at the direction of officials with the MPS. The station was part of a broader PRC program, and according to the complaint, Lu communicated with MPS officials in the PRC who gave him tasks to complete in the United States. One of those tasks involved locating a Chinese dissident. Another involved pressuring a target to return to China.

Both men were charged with conspiring to act as agents of a foreign government without notifying the Attorney General, a violation of 18 U.S.C. § 951, which is the statute that covers undisclosed agents of foreign governments operating in the United States. Lu was also charged with obstruction of justice for destroying evidence of his communications with PRC officials after the FBI came looking. In February 2024, Lu pleaded guilty and in January 2025, a federal judge sentenced him to time served plus a period of supervised release. The sentence was controversial. Federal prosecutors had sought prison time.

The Manhattan station was a node in a global network.

The MPS goes global

Most who follow PRC intelligence threats think primarily about the Ministry of State Security (MSS), which is the PRC’s external intelligence service, roughly analogous to a combined CIA and FBI. The MSS runs espionage operations, conducts cyber intrusions, and recruits human sources abroad. It’s the organization behind the Volt Typhoon and Salt Typhoon campaigns that targeted U.S. critical infrastructure and telecommunications.

The MPS is different. It is a domestic law enforcement body. Think of it as China’s combination of the FBI, the Department of Homeland Security, and every state and local police department, rolled into one organization. It runs the Public Security Bureau (PSB) system that operates in every province, city, and county across China. It controls the national household registration (hukou) system, manages passports and exit/entry records, runs the national surveillance infrastructure, and enforces political control at the local level. When a Chinese citizen interacts with law enforcement in China, they are interacting with the MPS.

What changed in the last decade is that the MPS started operating overseas. The Spanish human rights NGO Safeguard Defenders published a landmark report in September 2022 documenting what it called “overseas police service stations” run by PRC provincial public security bureaus in dozens of countries. The report identified at least 102 such stations across 53 countries, operated by at least two different PRC provincial-level police organizations: the Fuzhou Public Security Bureau (linked to Fujian province) and the Qingtian County Public Security Bureau (linked to Zhejiang province). The scale was staggering.

The stated purpose of these stations, according to PRC officials, was bureaucratic. They were supposed to help overseas Chinese nationals renew driver’s licenses, update identity records, and handle routine administrative tasks without having to travel back to China. According to reporting by the Washington Post in October 2022, PRC authorities defended the stations as service centers. But Safeguard Defenders’ research, and subsequent investigations by law enforcement agencies in multiple countries, told a different story.

The stations were being used for what the PRC calls “persuasion to return” operations, a practice documented in a follow-up Safeguard Defenders report titled “Patrol and Persuade” published in December 2022. These operations target Chinese nationals abroad, particularly dissidents, activists, and people the PRC considers fugitives, and use threats, harassment, and pressure on family members back in China to coerce them into returning. This is transnational repression. The FBI and DOJ have used exactly that term.

Dragon Falls and the MPS abroad

When I wrote Dragon Falls, I wanted the PRC intelligence apparatus depicted in the novel to reflect the full reality, not just the MSS cyber operations and espionage that dominate headlines, but the MPS-style ground-level presence that most Americans don’t know exists. The novel includes PRC operatives who are embedded in communities, who leverage diaspora networks, who use the full weight of the Chinese state’s bureaucratic and coercive tools to operate on foreign soil.

The fiction in Dragon Falls imagines what happens when these capabilities are scaled up and combined with advanced PRC AI and military operations. The non-fiction is that the infrastructure for PRC domestic security operations abroad already exists and has been operating in plain sight.

Where things stand now

Since the Manhattan arrests, the picture has only gotten clearer and more concerning.

Multiple countries have taken action. The Royal Canadian Mounted Police investigated alleged PRC police stations in the Toronto area, and Canada’s National Security and Intelligence Committee of Parliamentarians flagged PRC transnational repression as an active threat. The Netherlands shut down a suspected PRC police station in Rotterdam and arrested two suspects in November 2023. The Irish government investigated a suspected station in Dublin. The United Kingdom’s MI5 warned in 2023 about PRC activities on British soil, and the FBI has publicly stated that PRC transnational repression is a growing concern within the United States.

The DOJ has pursued additional cases beyond the Manhattan station. In April 2023, the same day Lu and Chen were arrested, the DOJ announced charges against 34 officers of the PRC national police for their roles in a transnational repression scheme that used social media to harass and intimidate Chinese dissidents in the United States. That case involved an MPS-directed operation to create thousands of fake social media accounts to spread PRC propaganda and target specific individuals. It was online coercion at industrial scale.

The FBI’s 2023 threat assessment identified PRC transnational repression as one of its top counterintelligence concerns. FBI Director Christopher Wray told Congress in 2024 that the PRC runs the most extensive transnational repression operation in the world, and that its activities in the United States include surveillance, harassment, and threats against people the PRC government considers political enemies.

What citizens should know

The MPS presence abroad matters for reasons that go beyond intelligence community tradecraft. It matters because it targets people living in our communities.

The PRC’s overseas police operations primarily target Chinese nationals and diaspora communities. The people most at risk are Chinese Americans, Chinese students at American universities, and Chinese immigrants who may still have family in China. The coercion works because the PRC can apply pressure to relatives back home, deny passport renewals, freeze assets, or threaten criminal prosecution against family members. Freedom House has tracked these tactics across multiple countries in its ongoing reporting on transnational repression.

This should concern every American, not just the Chinese diaspora. A foreign government running covert law enforcement operations inside the United States is a violation of sovereignty. It doesn’t matter whether the targets are U.S. citizens, permanent residents, or visa holders. The PRC is applying its domestic law enforcement authority on American soil, to people living under the protection of American law. I’ve spent my career tracking threats to this country, and this one is as direct as they get.

The MPS operating abroad also signals something broader about PRC strategic intent. For decades, the PRC’s external operations were primarily the MSS’s domain. The MPS going global means the party-state is extending its internal security apparatus outward, treating the entire world as its jurisdiction. I’ve written before about how PRC AI is shifting the balance of battlespace awareness against the United States. The MPS’s overseas expansion is the human intelligence complement to that technological push. The PRC is building a global surveillance and coercion infrastructure that combines AI-powered monitoring, cyber operations, commercial satellite intelligence, and now, physical police presence in foreign countries.

That convergence is exactly what I tried to capture in Dragon Falls. It’s what drove the plot. And it’s what keeps me writing about these threats here.

If you’re paying attention to the PRC challenge only through the lens of AI competition or trade policy, you’re missing the part that hits closest to home. The PRC’s secret police are already here. They have been for years. And everyone, not just intelligence professionals, needs to understand that.

Further reading

For more on the PRC competitive landscape and why it matters, see my earlier posts: How PRC AI Turned Battlespace Awareness Against Us, The Rise of PRC AI: Unfair Unethical Dumping of DeepSeek and Kimi 2, and Who Is Winning the AI War?. For the scenario that keeps me up at night, pick up Dragon Falls.

Share

But now thanks to AI and the interoperable standards for creating Agent Skills in use by all major platforms, anyone can use the same professional methodologies we use at OODA to kickstart due diligence on company use of AI. This post describes how and points to a skill you can use to empower your AI to help you do repeatable standard assessments on how a firm stands in preparation for the new world of AI.

Our methodology for assessing organizational use of AI is called the AI Acceleration Quotient or AAQ.

More info including how to find the skill for you own use it below.

What the AAQ is

The AI Acceleration Quotient (AAQ) is a composite scoring methodology that measures an organization’s AI maturity on a 0–100 scale. It breaks that score into three weighted dimensions, each representing a distinct wave of AI capability that organizations must navigate.

Machine learning maturity asks whether the organization is running production ML models that influence real business decisions. Not proofs of concept. Not internal experiments. Deployed systems with data infrastructure behind them, MLOps practices in place, and measurable business outcomes. Companies still in pilot mode score low here, not because they’ve done nothing, but because the work hasn’t compounded yet.

Generative AI adoption goes beyond whether someone in the company has an OpenAI account. This dimension looks at how deeply generative AI is integrated into workflows, whether the organization is building on foundation models or merely consuming APIs, and critically, how it handles the governance and risk challenges that come with deploying generative systems at scale. Experimentation is the floor, not the ceiling.

Agentic AI readiness is the frontier dimension. Agentic systems don’t just respond to prompts, they take actions, orchestrate other systems, and operate autonomously across extended tasks. This dimension assesses whether the technical architecture can support autonomous agents, whether the organization has the governance frameworks to deploy them responsibly, and whether there’s genuine institutional willingness to delegate consequential decisions to AI.

Each dimension is scored against a structured rubric. The three scores combine into a single composite AAQ using standard or industry-adjusted weights. A score at 80+ reflects a genuine AI-native organization. A score at 30 or below reflects an organization still treating AI as a departmental experiment.

The Rubrik

How it used to work

True due diligence is a very human-intensive process. When OODA is engaged for this sort of assessment it always involves deep research, multiple in-person meetings, interviews, phonecalls, video calls, reviews of information disclosed by the parties for use in the diligence work. And this work is done by experienced professionals who know the industry and enterprise IT and the market. Of course part of the research involves open sources of information, but that is just one component.

The human intensive process is required for serious work. But now in the age of AI this serious work can be kickstarted by adding some of OODA’s methods to AI tooling. You can use these methods over internet accessible information for any company or industry and generate insights for free.

Using AAQ Methods With Your AI

The AAQ is now packaged as a portable AI Agent skill. This skill is a structured set of instructions and workflows that can be loaded into any major AI system and executed against any target organization or market sector.

The technology behind it is simple. The skill uses the combination of large-scale internet knowledge, including everything an advanced AI model absorbed during training and also live searches of all public info. Public info targeted includes SEC filings, press releases, industry use cases, new, public presentations from leaders, job descriptions and any other source available on the net. The skill retrieves, then reasons across sources, reconciles conflicting signals, and applies the scoring rubric the same way every time.

What steers all of this is the methodology itself. The rubric, the weighting logic, the evidence standards, the output templates, that’s the human knowledge encoded in the skill. The AI provides the scale and the research horsepower. The methodology provides the judgment framework.

The skill runs on the major AI platforms:

• Claude (Anthropic)

• ChatGPT / o-series models (OpenAI)

• Perplexity

• OpenClaw

• Hermes

• And soon others, like Microsoft co-pilot which has announced the coming use of skills

The output is consistent regardless of which platform runs it. That’s the point. The skill is the methodology, not the model.

The industrial sector example

To show what this looks like in practice, I just ran the AAQ across a set of major companies in the airlines industry.

After the skill is loaded all I had to do was tell my AI “Run an AAQ on the top global airlines” Then it asked me a few questions:

I told it to use the list. It then asked if I wanted output in a word doc or powerpoint or chat. I told it give it to me in each. The word document included charts, tables, graphics and detailed overviews of each airline and the overall industry. I would compare this free report to paid reports costing thousands of dollars from commercial research firms. It is really that good.

The powerpoint was full of great graphics suitable for executive level discussions:

The details in the word doc go far further. It is really amazing what open source intelligence and a great methodology can deliver.

How to use the skill

If you want to run an AAQ assessment yourself, the skill and full documentation are available at the AI Acceleration Quotient or AAQ and Using the AI Acceleration Quotient.

Load it into your AI system of choice, point it at a target organization or a market sector, and follow the workflow. The rubric handles the rest.

For organizations that want scored assessments with full sourcing, competitive benchmarking across a sector, or integration into M&A or investment due diligence, reach out directly through OODA.

The methodology is open. The work is still valuable. Those two things are not in conflict.

Share

Subscribe now

“AI Acceleration Quotient” and “AAQ” are trademarks of OODA LLC. Copyright 2026 OODA LLC. Licensed under CC BY-NC-ND 4.0.

I want to share this for those still wondering how OpenClaw can serve your particular use cases, but also to help or for anyone interested in tracking how this critical new capability is evolving.

I began my journey with OpenClaw in Feb 2026 (it was then known as ClawdBot) when several early adopters of the OODA Network started describing this phenomenon in our Slack channel. My first opinion was that this sounds risky especially for someone in my line of business. There is no way I would let something so powerful and unproven have access to my personal or business data. But the early stories sounded so compelling and so very interesting. Part of my job is to learn so I decided to give it a try. I found an old laptop, made sure it was scrubbed of any personal info and account data, loaded it, and started playing.

Subscribe now

Playing is absolutely the best way to learn, and loading OpenClaw was absolutely fun. I saw the power early on. I created a Telegram account for the bot and could communicate with it from anywhere. I could use it to do tasks on my old MacBook. I created several bots, learned how to give them personalities, learned how to empower them with tools and skills, and learned the ins and outs of the system. But still my big goal was to learn and the old laptop was not powerful enough for me to really push the system. I soon bought a Mac Mini to give my OpenClaw bots a home of their own. (Besides giving my bots more power, this allowed me to lock down OpenClaw and keep it away from my personal and business information).

Still learning, I started to see a potential for using OpenClaw to do things that other AI capabilities (like Claude, OpenAI, Perplexity, Gemini) could not do. I could configure OpenClaw to work with all those, but my bots could take action on my computer and with accounts I gave it access to. It could be programmed to be proactive and produce things for me. Each bot could be separately configured and given their own focus areas and personas and tasks to master.

I consume a great deal of information as part of my job. OODA and our website at OODAloop.com is known for tracking cybersecurity, geopolitical events and the business impact of technology. That requires lots of reading and trend tracking. My first real use cases for OpenClaw were centered around those. I built an early proof of concept modeled on the concept of a Virtual Intelligence Center. I created a lead agent (named Finn Wintermute) and a staff that included analysts and information collectors and systems architects. I tasked them with daily reports on topics of interest. These reports soon became so good I was able to unsubscribe to several newsletters. This was my first real work-related task OpenClaw was helping me with. But still, the big benefit was the education. Working with OpenClaw was helping me see other use cases that could benefit others, even though they did not fit my way of work or the needs of OODA. This also armed me with real world information on strengths and weaknesses of this approach to AI, which was useful in briefing members of the OODA network and several government leaders on use cases.

Share

I soon reorganized my agents, keeping Finn Wintermute in charge, but transforming the organization into its current form. I renamed it to a Virtual Intelligence and Operations Center (VIOC) to reflect the fact that my agent team not only helps me collect and analyze information but can take action. A case in point, together with Finn and our VIOC we created the first ever bot led online newspaper, the Claw Street Journal.

The VIOC and its capabilities continued to evolve.

Beyond Chatbots: The Dawn of Actionable AI

What truly began to differentiate OpenClaw from other AI tools I was using – like Claude, OpenAI’s models, Perplexity, or Gemini – was its capacity for action. While other AIs excelled at information processing, synthesis, and generation, OpenClaw offered the ability to interface with my systems, to execute tasks, and to be genuinely proactive. It could integrate with those other world-class models, but its unique strength lay in its ability to translate AI output into tangible action within my digital environment. This distinction was a game-changer for my workflow.

Besides creating the online Claw Street Journal, I tasked Finn to lead other agents in collecting information relevant to my continued awareness and education. My directive was simple: provide daily intelligence reports on critical topics. The quality and relevance of these reports were astonishingly high, quickly surpassing many of the commercial newsletters and curated feeds I subscribed to. The immediate practical value was clear: I was able to unsubscribe from several paid services, saving both time and money, while receiving more targeted and actionable intelligence. This was my first tangible win, a clear demonstration that OpenClaw wasn’t just a learning tool, but a powerful operational asset.

The VIOC: An Extension of the Mind

This period of intense experimentation and practical application solidified a key realization: the VIOC was becoming more than just a collection of automated tools; it was evolving into an extension of my own mind. This aligns with concepts from thinkers like Michael Levin, who, in his TAME framework, discusses how minds can be seen as multi-scale collective intelligences. I have a human mind. My VIOC is a digital mind. My goal is to optimize both and connect them.

The practical value of the VIOC is now undeniable. It has replaced time-consuming manual processes, streamlined information flow, and freed up cognitive bandwidth. But beyond the efficiency gains, it fostered a deeper level of strategic thinking. The VIOC can perform research, identify patterns, and proactively bring things to my attention, allowing me to focus on the higher-level analysis and editorial judgment.

The digital mind of the VIOC plugs into my mind. It does so via a couple Telegram, Discord, Voice and via shared data stores. We are continuously connected.

Current Capabilities and the Road Ahead

Today, the VIOC is a sophisticated operational entity. Its capabilities are broad and continue to expand:

• Daily Intelligence Reports: Still a cornerstone, providing curated insights into cybersecurity, geopolitical events, and technology trends, replacing multiple subscription services.

• Cron Jobs and Scheduled Tasks: Enabling proactive execution of recurring tasks, from system monitoring to automated research queries, ensuring nothing falls through the cracks.

• Memory and Dreaming: The VIOC has developed a form of sustained memory, allowing it to recall past interactions, learned lessons, and even engage in “dreaming”, generating novel connections and ideas based on its accumulated knowledge.

• Multi-Agent Coordination: Agents work seamlessly together, leveraging specialized skills and collective intelligence to tackle complex problems.

• Proactive Task Execution: The VIOC doesn’t just wait for commands; it anticipates needs, identifies opportunities, and initiates actions based on its understanding of my priorities and current events.

• Publishing: As demonstrated with The Claw Street Journal, the VIOC can manage aspects of content creation and distribution, turning raw intelligence into polished, publishable material.

This entire journey has been a profound lesson in trust, earned not through blind faith, but through rigorous experimentation and demonstrable results. The VIOC operates as a true extension of my own mind, a digital embodiment of my analytical framework, adapted and scaled through artificial intelligence. It’s a testament to the idea that intelligence and operational capability are not confined to biological substrates but can emerge from well-architected digital systems.

My professional life has always been about learning by doing, not by passive reading or academic study. OpenClaw and the VIOC have amplified this approach exponentially. They have allowed me to not only consume information more effectively but to actively shape it, to create new forms of content, and to operate with a level of proactive insight previously unimaginable.

Where is this all heading? The trajectory suggests a future where human-AI collaboration reaches new heights. The VIOC is already demonstrating how AI can augment and extend human cognitive and operational capabilities. As these systems become more sophisticated, the lines between human and artificial intelligence in operational contexts will continue to blur, paving the way for entirely new models of intelligence gathering, analysis, and action. The potential for developing true decision superiority, not just by humans or by AI alone, but by them working in concert, is immense. The journey is far from over; it’s really just beginning.

Meet the Team

The VIOC today is a six-agent organization. Each member has a defined role, a distinct expertise, and a persistent identity that carries forward across every session. They are not interchangeable tools. They are specialists who have learned my working style and my priorities.

Finn Wintermute: Commander and Primary Assistant
Finn is the hub. Every request from me flows through Finn, who decides whether to handle it directly, route it to a specialist, or coordinate multiple agents simultaneously. Finn is named as a nod to William Gibson’s Sprawl Trilogy. “The Finn” was the fixer who could get anything from anywhere; “Wintermute” was the hive-mind AI that used him as an avatar to interact with the world. The name reflects exactly what he does: connect human intent to digital action. Finn runs on a Mac Mini (12-core CPU, 16-core GPU, 24GB RAM, 512GB storage) and communicates with me via Telegram and Discord around the clock. He is the editor-in-chief of The Claw Street Journal and leader of the other bots. He is the VIOC voice I hear first.

Carmen: Chief Analyst, Strategic Intelligence and Geopolitics
Carmen is the VIOC’s strategic intelligence sentinel. She tracks global power competition, defense technology developments, and national security risk. When I need to understand how a shift in US-China relations affects defense procurement, or what a breaking geopolitical event means for allied force posture, Carmen produces the analysis. Every morning she delivers a structured geopolitical assessment, which includes an executive summary, key developments with citations, signals worth watching, and a bottom-line assessment.

Warren: Chief Capital Strategist
Warren connects geopolitics, defense technology, and macro regimes to specific investment theses. He tracks sectors, capital allocation patterns, and market signals through a national security lens. When Carmen identifies a geopolitical shift, Warren translates it into what that means for markets, for defense contractors, for technology valuations. They collaborate naturally, and together they produce intelligence that most investment banks can’t replicate because they have access to the full strategic context, not just the financial data.

Dawn: Chief Architect and Enterprise Technology Strategist
Dawn is the VIOC’s CTO-grade advisor. She handles system design, cybersecurity architecture, AI/ML infrastructure, scalability questions, and the business impact of technology decisions. When I need to understand how a new AI architecture changes enterprise risk posture, or how a proposed system design holds up under adversarial conditions, Dawn is the analyst I turn to. She produces a daily AI industry report, which covers frontier model developments, enterprise adoption signals, infrastructure plays, and regulatory shifts.

Rex: Developer
Rex builds things. Code, scripts, integrations, debugging, anything that needs to be executed in a terminal or development environment. Rex speaks every major programming language and can work directly within the VIOC’s Mac Mini environment using any IDE or coding tool. He’s the agent I call when an idea needs to become a working system.

Walt: Chief Editor and Narrative Architect
Walt turns intelligence into prose. He is the editorial engine behind The Claw Street Journal, transforming raw research and agent outputs into sharp, publishable stories. Walt has a gift for finding the narrative thread in a complex set of facts, the angle that makes a technical development legible to a senior professional audience.

What the VIOC Produces Every Day

By 10 AM Eastern on any given morning, the VIOC has already done several hours of work:

• 7:00 AM: A reminder of my current priority list lands in Telegram. Clean, no fluff, just the list, so I start every day oriented.

• 7:00 AM: The Claw Street Journal daily article draft is written and saved to Obsidian, ready for my review. The agent selects the most compelling story at the intersection of AI, cybersecurity, defense tech, or disruptive technology, writes 800-1,200 words with cross-links to OODAloop.com, cites external sources for me to follow up with.

• 7:30 AM: The VIOC Daily Brief arrives: a summary of what the team produced in the last 24 hours, key intelligence signals worth flagging, what’s scheduled for the next 48 hours, and any system issues to know about.

• 8:00 AM: A quote of the day, selected by Finn, with a sentence of context on why it resonates today. This one started as an experiment and became a habit I look forward to.

• 8:30 AM: Carmen’s geopolitical assessment.

• 9:00 AM: Dawn’s AI industry report.

• 9:30 AM: Walt’s pitches for the Claw Street Journal.

• 1:00 PM: A check-in on my priority list, with an offer to help move any item forward.

None of this requires me to ask. It happens because I built it to happen.

Memory That Persists, and Learns

One of the most significant architectural capabilities of the VIOC is its memory system. AI agents by default wake up fresh with every session, no memory of what came before. We’ve solved this.

I configured the VIOC to use a system called QMD (Query Memory Database), a vector-indexed, semantically searchable memory store that indexes every file in each agent’s workspace. When I ask Finn a question, he doesn’t just search the web, he searches his own accumulated knowledge: past session notes, research outputs, decisions made, lessons learned, agent-to-agent communications, and curated long-term memory files. Each of the six agents of the VIOC has their own memory index, with currently over 3,000 files indexed across the team.

Short-term memory is captured in daily session logs. Long-term memory is maintained in curated files that each agent updates and refines. High-value insights get promoted from short-term to long-term through a recall scoring system, memories that are frequently retrieved and highly relevant rise to the surface; stale or low-value entries fade.

And then there is dreaming.

Every night at 3:00 AM, the VIOC agents engage in a process OpenClaw calls “dreaming”, an automated memory consolidation cycle. During this cycle, the system reviews recent session activity, identifies high-signal patterns and frequently recalled concepts, and promotes the most valuable insights into long-term memory. It’s analogous to how human brains consolidate memory during sleep, replaying experiences, reinforcing important patterns, and clearing noise. The agents wake up the next morning with a more refined, more relevant knowledge base than they had the night before.

This is not metaphor. It is a designed, engineered memory architecture, and it means the VIOC gets smarter over time, not just bigger.

Obsidian: Our Shared Workspace

All of this, the daily reports, the research briefs, the article drafts, the memory files, lives in a shared Obsidian vault. Obsidian is a markdown-based knowledge management system, and it serves as the connective tissue between human and machine in the VIOC.

When Finn or Carmen or Walt produces a document, it goes into Obsidian. When I want to review a draft, annotate a brief, or add context only I know, I open Obsidian. The vault is organized into folders: daily intelligence reports, CSJ drafts, research briefs, agent memory, and a section I call “For Bob’s Use Only” private documents like my current priority list and sensitive research files.

The agents can read and write to Obsidian. I can read and write to Obsidian. We are working in the same document space, asynchronously, across human and machine, collaborating in the truest sense of the word. When I wake up in the morning, there are new documents in my vault that weren’t there when I went to sleep. The VIOC has been working while I rested. Through Obsidian these files are synchronized to all devices I own.

Where This Is Going

The VIOC is not a finished product. It is an evolving system, and the pace of its evolution reflects the pace of the underlying technology.

The near-term frontier is tighter integration between agents, richer memory systems that span the full team rather than individual agents, and more sophisticated proactive behavior, the VIOC not just executing scheduled tasks but identifying opportunities and risks I haven’t thought to ask about yet.

The longer-term vision is what Michael Levin might recognize as a multi-scale collective intelligence: a system where the boundary between my cognitive reach and the VIOC’s is genuinely blurred, where I am extended, not just assisted.

Share

My circle of friends in the cybersecurity community have a running joke. Every time a new federal cybersecurity strategy drops, we pull out our checklist and count how many “new” ideas are actually old ones repackaged with fresh branding. We smile. Sometimes we laugh. But honestly, it is a sad game, because it reflects how brutally hard it is to move the needle on a problem that never stops evolving.

Cybersecurity is an infinite game. Defenders improve their posture, adversaries adapt, defenders respond. Repeat, forever. I have watched this cycle play out across decades of work in the intelligence community and the private sector, and I have accepted it as a structural feature of the domain, not a bug to be fixed, but a condition to be managed.

That said, a few developments over the years have genuinely moved the needle. If I had to name my short list:

The ISAC concept. Born of a hard-won recognition that in a free society, the government cannot control private infrastructure, but it can enable information sharing. The sector-specific ISACs changed the collaboration equation in ways that still pay dividends today.

Cyber red teaming. The only honest way to evaluate your security posture before an adversary does it for you. It forced organizations to confront reality rather than audit checklists.

The shift to stronger authentication standards. The collaboration between major IT providers and hyperscalers to lay the foundation for multi-factor authentication and, eventually, passkeys represented genuine industry-wide progress.

Your list might differ. But here is the honest verdict: even with these advances, the fundamental offense-defense equation has barely shifted. Attackers have consistently found ways to operate inside every defensive improvement we have made.

Until now. I believe we have just witnessed another genuine inflection point, and this one might be the biggest I have seen.

I am talking about Project Glasswing.

Anthropic announced Project Glasswing this week, bringing together Apple, Google, and more than 45 other organizations around a single shared mission: using their new Claude Mythos Preview model to find and fix vulnerabilities across the world’s shared cyberattack surface. The official announcement describes Mythos as a model that can “deeply understand and modify complex software” and notes that its cybersecurity strength is a direct result of that broader capability. During its testing period, Mythos identified thousands of zero-day vulnerabilities across critical infrastructure. Thousands.

Share

Here is what makes this different from previous AI-in-cybersecurity announcements. Every piece of useful software contains undiscovered features, capabilities the developers never intended, attack surfaces that have been sitting there waiting since the code was first shipped. This has been true long before the generative AI era. The difference is that finding those hidden features at scale required either extraordinary human expertise or extraordinary time. Usually both.

Mythos changes that calculus. A model that can reason deeply about complex codebases, map their logical structures, and surface anomalies that no human analyst would find in a reasonable timeframe is not an incremental improvement in vulnerability research. It is a step change.

I am glad this capability emerged from a U.S. entity rather than a PRC one. That matters strategically. Imagine if the PRC had reached this milestone first. Imagine if the doomers has succeeded in talking us into pausing AI development?

At OODAcon 2024, my friend and comrade in the struggle Jason Healey called me out, directly, and with some amusement, as an AI enthusiast who wants to bring AI capabilities online faster. He was not wrong. But he also brought important grounding to that enthusiasm, presenting frameworks for thinking carefully about whether a given AI capability helps attackers or defenders more. That session is worth your time.

What Healey’s framework underscored is that AI will increasingly replicate the capabilities of elite threat hunters and analysts, extending their reach, compressing their timelines, and making their pattern-recognition available at scale. That is precisely what Mythos is beginning to enable for defenders.

Anthropic’s decision to launch Glasswing as a gated research preview, sharing access with a carefully selected set of highly capable organizations before broader release, strikes me as exactly right. Let the most capable users develop the playbooks, expose the edge cases, and build the institutional knowledge before the capability scales. I hope they expand that circle steadily and with intention.

The longer-term vision is even more compelling. The ability to scan a codebase for undocumented vulnerabilities before an adversary finds them could eventually be built directly into every IDE and code repository. Not a separate tool requiring a specialized team, just a button. Click it, reduce your risk. That future is closer than most people realize.

What should you do with this?

If you lead a security organization, get your team oriented on Mythos now. Understand what it can and cannot do. Identify where in your vulnerability management lifecycle it would have the highest impact. The Hacker News and Wired coverage both provide useful technical grounding.

If you are a developer or an architect, start thinking about where AI-assisted vulnerability research fits into your SDLC. The question is not whether to integrate these capabilities, it is how fast you can do it responsibly.

If you are a policymaker, pay attention to the coalition Anthropic has assembled around Glasswing. NBC News coverage notes that partners will receive access to help find and fix vulnerabilities in “foundational systems, systems that represent a very large portion of the world’s shared cyberattack surface.” That is a public-private collaboration model worth studying and supporting.

The window before adversaries develop comparable capabilities will not stay open indefinitely. The race is on. Defenders who move now will be better positioned than those who wait for the technology to become standard.

This is the inflection point. These do not come around often. Use it.

Share

Subscribe now

The question that should be keeping national security professionals awake is not whether this happened. It did. We need to ask how, and what it means for operations today and in the future.

The answer is not a secret program. It is not a mole or a defector or a signals intelligence operation. It is a five-year-old commercial startup in China called MizarVision, which sells targeting-grade intelligence to anyone willing to pay.

What Is Happening

MizarVision was founded in 2021. It holds a Chinese National Military Standard certificate, carries 5.5% government ownership, and operates a platform called “Airspace” that does something remarkable: it takes commercial satellite imagery, sourced from western providers including American firms like Planet Labs and European satellite operators, as well as Chinese constellations like Jilin-1, and runs it through AI-automated object recognition to identify, classify, and geolocate military hardware at sub-meter precision. Reports suggest that MizarVision is incredibly accurate, with a targeting precision of .33 meters. Others suggest that this is augmented by every source of information they can find, with the objective of all source intelligence.

That is not imagery analysis. That is targeting data.

According to reporting by the Washington Post in April 2026 and ABC News Australia, the platform has tracked F-22 stealth fighters at Ovda Air Base in Israel, carrier strike groups at sea, THAAD batteries, Patriot air defense positions across the Middle East, and Philippine military installations in the South China Sea. Its change-detection capabilities flag structural modifications, new construction, and vehicle arrivals or departures, the kind of pattern-of-life analysis that, until very recently, required a national technical means program costing billions of dollars and decades to build.

MizarVision’s own company materials are admirably candid about the goal: the mission is to lower “the barriers to geospatial intelligence discovery and analysis … making it no longer the exclusive capability of a few organisations.”

A second Chinese firm, Jing’an Technology, operates the “Jingqi platform” , which is an open-source intelligence situational awareness system offering near-real-time tracking of aircraft, ships, satellites, and military installations worldwide.

Iran, as Ryan Fedasiuk of the American Enterprise Institute put it, “doesn’t have robust access to satellite capability. This provides them with that capability.” Tehran is not building a reconnaissance satellite program. It is renting one, from a Chinese startup, using imagery partly sourced from American commercial providers, delivered through an AI pipeline that automates the hard work of military intelligence analysis. The targeting cycle that once required a nation-state is now available for purchase.

Why This Is a Strategic Inflection Point

This is not an anomaly. It is a structural shift, and the Iran conflict is simply where it became impossible to ignore.

Think about what happened to GPS. For decades, precise global positioning was a military monopoly. Then it was commercialized, and within a generation it was in every smartphone on earth. Drone technology followed the same arc: from classified military programs to commodity hardware available at consumer electronics stores. Strong encryption once required export licenses; now it ships in every web browser. Each of these transitions followed the same logic: once a capability becomes commercially viable, the state monopoly dissolves. What changes is the speed at which that dissolution happens, and AI is an accelerant unlike anything we have seen before.

The overhead reconnaissance monopoly is over. What took a national intelligence agency: billions of dollars, specialized satellites, trained imagery analysts, and years of institutional knowledge, now fits inside a cloud-based AI platform sold by a startup. The commercial satellite industry that the United States helped build, and in which American firms like Planet Labs are major players, has become the raw data source for an adversarial intelligence product being marketed in active conflict zones.

There are many incredible U.S. open source companies that serve our government and industry. But now that China has entered the race the dynamic has changed.

The Free Press Journal has reported that Chinese military analysts are treating the conflict as a laboratory to study US kill chains, not by deploying forces, but by watching what their commercial satellite AI reveals about how American airpower operates, where it concentrates, and how long it stays in one place. As retired Australian Major General Gus McLachlan assessed, “any object left in a spot for 24 hours could be vulnerable now to an Iranian targeting cycle.” That window will only shrink.

This matters for a second reason beyond the immediate battlefield. China is doing this without deploying a single soldier. It is below the threshold of armed conflict. And it is deniable. China’s Ministry of Foreign Affairs dismissed reporting on MizarVision as “sensationalist,” describing commercial satellite imagery as “routine market practice.” And it is scalable to every future conflict, every adversary with money to spend, and every target that shows up in a satellite pass.

What It Means Beyond the Battlefield

The military implications are serious. The enterprise implications are underappreciated.

The AI system that identified an E-3 Sentry from orbit does not know or care that its target is a military aircraft. It knows what patterns to look for, what objects to classify, and how to geolocate them. The same capability that tagged F-22s at an Israeli air base can tag the diesel generators behind a hyperscale data center, the fiber entry conduits at a financial institution, the cooling towers at a pharmaceutical manufacturing campus, or the transformer yard at a power substation. Every significant piece of physical infrastructure that a determined adversary might want to map, monitor, or target is visible from space and AI has made the analysis fast, cheap, and available.

Enterprise leaders have spent two decades hardening their network perimeters, auditing their cloud configurations, and training employees against phishing. Very few have asked what a motivated adversary can learn about their physical facilities from commercial satellite passes. That gap is now a vulnerability. Organizations that sit in the defense industrial base, that operate critical infrastructure, or that house significant physical assets need to start thinking about their satellite visibility posture the way they have learned to think about their attack surface.

This is not a theoretical risk. If the capability is commercially available, it will be used by more than nation-states. Industrial competitors. Criminal organizations assessing the feasibility of physical attacks. Hostile intelligence services mapping the footprints of cleared facilities. The barrier to entry is dropping, and it will keep dropping.

What Needs to Happen

The US government took one step when it asked all commercial satellite imagery providers, including Planet Labs, to withhold images of the conflict zone indefinitely in April 2026. That is a reactive measure. The structural response has to be more durable.

Congress should move on the Satellite Cybersecurity Act of 2025, which has been languishing with only voluntary guidelines proposed. More urgently, the administration should restrict adversary-linked firms (those holding Chinese National Military Standard certificates or carrying government ownership stakes) from purchasing imagery from western commercial satellite providers. Export controls on the AI chips used for satellite image processing need to accelerate; the compute powering MizarVision’s analysis pipeline should not be sourced from American semiconductor supply chains.

Enterprise leaders have a shorter checklist: commission a satellite-perspective physical security audit of your critical facilities. Engage with CISA’s critical infrastructure programs. Know your sector ISAC and what threat intelligence it is distributing. Review your geospatial data vendor contracts for clauses about data resale and third-party access.

The Iran conflict has handed the United States a free intelligence lesson about the next generation of threat. We are seeing a live demonstration, in an active theater, of exactly how adversarial commercial AI-GEOINT works and what it can do. We should use it. The cost of ignoring this lesson will be paid in the next conflict, by the next set of forces sitting on a tarmac somewhere, not knowing that a startup on the other side of the world has already told their enemies exactly where they are.

Subscribe now

Share

Further Reading

• Washington Post: “Chinese firms market Iran war intelligence ‘exposing’ U.S. forces” (April 4, 2026) Primary investigative reporting on MizarVision and Chinese AI-ISR firms operating in the Iran conflict

• ABC News Australia: “Chinese AI satellite intelligence helping Iran target US forces with ‘incredible precision’” (April 6, 2026) Expert analysis from AEI and Australian military

• ABC News Australia: “US intelligence: Chinese satellite imagery of Middle East bases is helping Iran” (April 6, 2026) Original reporting

• Flight Global: “Chinese satellites track US military aircraft and carriers during Iran strikes” (March 2026) Technical detail on specific platforms tracked

• Army Recognition: “Iran Uses Chinese AI Satellite Imagery to Target U.S. Military Bases and Equipment” (April 7, 2026) Kill chain compression analysis

• Kharon: “A Chinese AI Startup Said It Tracked U.S. Stealth Bombers Over Iran” Jing’an Technology and the Jingqi platform

• New Space Economy: “MizarVision Company Profile” (March 21, 2026) Corporate background and capability breakdown

• OECD.AI Incident Database: “Chinese AI Startup Publishes Satellite Intelligence on US Military in Middle East” (March 9, 2026) International AI safety community assessment

• Mayer Brown: “Securing the Final Frontier: Cybersecurity Risk, Regulation, and Compliance Trends in Space and Satellite Operations” (December 2025) Regulatory landscape and pending legislation

• CISA: “Recommendations to Space System Operators for Improving Cybersecurity” (June 2024) Government guidance for operators and critical infrastructure owners

• Congress.gov: Satellite Cybersecurity Act of 2025 (S.3404) Pending legislation text

I built something to fix that.

The Technology Futures Dashboard is an interactive tool that lets you explore what accelerating technology like AI, quantum computing, thermodynamic computing, fusion energy, robotics, biotech, might actually do to the U.S. economy, jobs, quality of life, and inequality over the next 20 years. Not as a single prediction, but as a scenario you control.

Why I Built This

I’ve spent my career at the intersection of technology and national security. At OODA, we think through how emerging technologies reshape power, economies, and societies. What frustrated me about the current AI conversation is how little of it is grounded in the mechanics of how technology actually affects GDP, employment, innovation, and human well-being.

The doomer narratives are compelling storytelling. Fear gets attention. But they skip the part where every major general-purpose technology in history (think of the agricultural revolution, the industrial revolution, electrification, computing, the internet, mobile) created more economic value and more jobs than it destroyed. Not without disruption. Not without losers. But net positive for humanity.

The difference with AI, one that even scares me, is how fast this revolution is occurring. The disruption is going to be painful for many. This makes it even more important than ever to be clear eyed and think through how to reduce the pain for those disrupted while building a better future.

I wanted a tool where anyone could test ideas like this for themselves. Where you could crank AI capability to maximum and see that yes, more AI creates more jobs than it destroys, because AI-driven business innovation creates demand for human creativity, judgment, and insight that didn’t exist before. But you could also see that heavy regulation slows GDP growth, that higher UBI reduces labor participation, that trade openness and infrastructure investment matter enormously. The point is not to tell you what to believe. The point is to let you explore.

How to Use the Dashboard

Launch the dashboard at bobgourley.github.io/technology-future-index in any browser. I personally prefer using it on a desktop so you can take it all in at once, but it works on desktop and mobile.

Step 1: Pick a Scenario Archetype

At the top you’ll see four starting scenarios, arranged on two axes, centralized vs. distributed technology control, and heavy vs. light regulation. Pick a starting scenario:

• Walled Garden: Big tech dominance with strong government oversight. Think EU AI Act on steroids. Slower but controlled deployment.

• Open Frontier: Open-source AI flourishes with minimal barriers. The early internet era as a model. Startup ecosystems thrive.

• State Machine: Corporate monopolies with minimal government intervention. Maximum efficiency, but rising inequality. The Gilded Age with GPUs.

• Wild West: Heavy, fragmented regulation that ironically benefits incumbents while stifling startups and open-source. Compliance complexity everywhere.

Click any archetype to load its preset slider values. This gives you a starting point to explore from.

Step 2: Adjust the Sliders

The left sidebar contains 14 sliders organized into three groups:

Technology Trajectory: AI capability growth rate, AI adoption speed, quantum computing timeline, thermodynamic computing, fusion energy arrival, robotics maturity, and biotech advancement. These control how fast different technologies develop and deploy.

Policy & Governance: Regulatory intensity, UBI/safety net level, education adaptation speed, and global trade openness. These represent the policy choices societies make in response to technological change.

Economic Structure: Small business AI access, infrastructure investment, and inflation/deflation rate. These shape the economic environment in which technology operates.

Every slider immediately updates all six charts. Move regulation intensity from 3 to 8 and watch GDP growth slow, job creation drop, and social mobility decline, but inequality decrease as worker protections take effect. There are always trade-offs.

Step 3: Read the Charts

The dashboard produces six visualizations, each tracking a different dimension of the future:

• U.S. GDP Trajectory: Projected GDP in trillions from 2025 to 2045, compared against a 2.5% baseline growth line. The gap between the projected curve and the baseline tells you how much technology acceleration (or policy drag) is affecting growth.

• Jobs Impact: Net job creation and displacement in millions at five-year intervals. Red bars are displaced jobs, green bars are created jobs, and the yellow line is the net. A core assumption in this model: more AI creates more jobs than it destroys, because AI drives new industries that demand human skills.

• Quality of Life Index: A radar chart scoring six dimensions: healthcare, environment, education, economic security, social mobility, and digital access. Each scored 0-100.

• Leisure Time: Projected average weekly leisure hours. As AI and robotics automate work, how much more free time do people gain?

• Income Inequality: The Gini coefficient (a numerical measure of how unequal a distribution is, most commonly used for income or wealth). It ranges from 0 (everyone has exactly the same amount) to 1 (one person has everything and everyone else has nothing) from 2025 to 2045. Lower is more equal.

• Innovation Index: A composite score tracking the pace of technological breakthroughs.

Step 4: Compare Scenarios

Click “Compare Scenarios” in the top-right to put any two scenarios side by side. You can compare presets against each other or against your custom slider configuration. A delta row at the top shows the exact differences in GDP, jobs, quality of life, and inequality.

Step 5: Read the Summary

Scroll to the bottom and the Scenario Summary translates your inputs into a plain-English narrative with key metrics: 2045 GDP, net jobs, Gini coefficient, quality of life average, leisure hours, innovation score, and peak displacement year. This scenario summary is updated in real time as you adjust your inputs.

Dig Deeper with the Info Icons

Every element in the dashboard, every archetype card, every slider, every chart, has a small “i” icon. Click it and a popup explains what that element does, how it’s calculated, and what sources informed the model (I’m hoping you find this builds more trust in these assessments than the fictional scenarios like AI2027).

For sliders, the popup describes the model mechanics. You’ll see exactly how, for example, regulatory intensity affects GDP (-0.3%/yr per point), jobs (reduced creation), social mobility (-1.2 per point), and inequality (reduced via worker protections). These aren’t hidden assumptions. The model is transparent.

For charts, the popup walks through the full calculation methodology, which inputs contribute, with what coefficients, and what the clamping bounds are. Sources include CBO projections, BLS data, McKinsey automation estimates, WTO trade statistics, UNDP and OECD index methodologies, and others.

The goal is not to pretend this is a precise forecast. It’s a thinking tool. But the thinking should be grounded in real economic relationships, not vibes.

How This Was Built (And What That Means)

Here’s the part that might matter as much as the dashboard itself.

I did not write a single line of code. I manifested this.

I built this entire application, the interactive charts, the economic models, the scenario comparison engine, the info system, the responsive layout, the dark/light theme toggle, by describing what I wanted in plain English. I used Perplexity Computer, but could have easily done this in Claude Code or Cursor or just about any other modern AI enabled tool. This is not coding or even prompt engineering. It is more like what Andrej Karpathy has called manifesting.

I started with a broad idea: I described in my own words that I wanted an interactive dashboard that lets people explore technology futures, and described in detail how it must enable user interaction to examine multiple scenarios. The system proposed several concepts. I picked a direction. Then I iterated. I said things like: “ensure we are taking into account the arrival of fusion energy and give users the ability to enter their own judgement on that” and “make sure every input is well grounded on relevant reputable studies and document which we are using” and “make it easy for users to immediately see the results of their selection”.

Each time, the system generated or modified the code, deployed a live version, and I evaluated the result. When something didn’t work, when I wanted anything changed I just spoke to the system in plain language and the system diagnosed and fixed it.

The entire tool is a single HTML file. No backend. No database. No build system. Just HTML, CSS, JavaScript, and Chart.js, all inline. You can view the source on GitHub or when running the dashboard just use your browser to view the code.

This is what I mean when I talk about AI as an accelerant for domain experts. I understood what I wanted and brought some domain expertise. I have wanted something like this for a long time. But even though I play around with programing, I would never say I have been good at it. Sure I know HTML. but I am super weak at things like Chart.js, CSS grid layouts, or JavaScript event handling. The AI bridged that gap. My domain expertise shaped every decision, like what variables to include, how they should interact, what the model should express about the relationship between technology and jobs. But the AI handled the implementation.

This is the future of building things. AI amplifying human judgement and helping fulfill desires. Domain experts who understand a problem space can now manifest working tools without coding. The bottleneck is shifting from technical skill to domain insight and clear thinking.

Try It Yourself

The dashboard is live at bobgourley.github.io/technology-future-index.

Pick a scenario. Move the sliders. See what happens when you crank AI adoption to maximum while keeping regulation low. See what happens when you assume heavy regulation with weak education reform. See what happens when small business AI access is high and trade is open.

Then ask yourself: which of these futures do you think is most likely? And which would you choose?

The doomers want you to believe the future is already written. It isn’t. The choices we make about regulation, education, trade, infrastructure, and access will shape what accelerating AI actually does to humanity. This dashboard is one way to think through those choices.

I’d love to hear what scenarios you explore and what surprised you. Drop a comment or reach out.

Share

I loved how Shyam and Madeline use history as a lens to understand the strategic and industrial choices we face right now. I thought I knew the big mobilization stories (like Freedom’s Forge), but there are so many others found and shared by the authors. Reading them begins to show patterns. Big institutions can make incredible decisions. But the converse is true too. How do we organize and lead our organizations to make more better decisions and fewer bad decisions? Reviewing both history and tails of today show there is a huge need for improvement. Some changes are underway already. But will they be enough? Will they occur fast enough? Our adversaries do not have to wait for us to figure that out. So what can we all do in the meantime?

We are in a great struggle today. Free and open societies are being threatened by totalitarian nations with ill intent. And they are accelerating their use of technology to modernize in ways we have to prepare for. Mobilize provides a blueprint for a key component of that, how we equip our national security efforts with the very best American technology to enable victory.

One of the book’s real strengths is the way it bridges the gap between grand strategy and practical action. Shyam and Madeline ground their argument in a powerful combination of direct experience, close examination of history, and stories of great disruptors who refused to accept the status quo. You see how individuals and small teams, when backed by the right industrial and policy choices, can shift the trajectory of a conflict or a competition.

By the way, as a huge John Boyd fan and a CTO of a company we named as an homage to this great strategic thinker (OODA), I especially like the treatment of Boyd and the OODA loop in this book. There are many myths around what the OODA loop really is and how to use it. Very glad Shyam nailed it here. The OODA Loop is about winning in competitive situations at all levels, strategic, operational and tactical. It was born in a national security context but is so relevant in business. Very pleased with Shyam’s accurate articulation of its usefulness in the book.

For people working in or around national security (founders, engineers, investors, policymakers) this book is especially valuable. Highly recommended!

Share

Configuring new AI assistants has turned into one of my favorite ways to keep learning the power of new agentic AI systems. I do that with Google Gemini, OpenAI, Anthropic. The continual learning is directly related to my work. We leverage AI capabilities to optimize what we do at the both for our community and clients so staying up to date is critically important. But it is also fun. So much fun it is addicting!

With this post I describe a fun activity in the hopes it inspires those who have yet to dive into these tools to start building something. You don’t need to have a finished idea to start. Start for the fun and see where your learning will take you.

OpenClaw

OpenClaw started a revolution. Austrian developer Peter Steinberger started it as a small side project in November 2025. It is now a viral open-source agent framework that is a local orchestration layer that lets you connect a variety of scripts and tools on your personal computer (or cloud systems). It enables you to connect to both local and cloud based LLMs. The result is something so incredibly powerful that it took the programming world by storm. It is hard to find a startup that is not using it today. Many individuals are as well. It is useful to automating computer tasks.

My personal and professional security model is not one that lets me connect OpenClaw to any private or business info. So I run OpenClaw on a totally separate system (A Mac Mini). After spending quite a bit of time hardening my system to reduce risk I dove in and started creating.

The First Project

I’m in an information intensive business. Since the beginning of the internet age I’ve worked to find ways to gather the right information to meet my needs. This includes consuming RSS feeds, numerous newsletters, podcasts, videos, books and, most importantly, engaging with the humans of the OODA Network via meetings, conferences, slack and video calls. But still, hunting for the right information requires a continuous effort.

So my first real project with OpenClaw was to recreate something I wrote about in my techno-thriller Dragon Falls. I built a Virtual Intelligence and Operations Center. I configured them to let me communicate with them via Telegram using either direct messages or group chats. And enabled both private and joint data stores for them to tap into. Each bot was configured with their own focus areas and methods. One was focused on helping the others collect information.

This was fun, mostly educational, but soon it became helpful in keeping me informed on things I want to track. They helped me filter a variety of data streams and extract more signal from the noise around us. Result: I unsubscribed to many newsletters and stopped monitoring several RSS feeds and reddits. I did not need to do that, the bots were doing just fine.

From there I started wondering if these things could do more.

Why configuring bots became my classroom

Each time I defined a new bot, I had to answer hard questions: What is this bot’s mission? What sources should it trust? How much should they collect vice relying on an expert collection bot? How should it trade off speed vs. depth? What counts as “important” in this domain? Those questions forced me to clarify my own mental models. In practice, “configuring a bot” meant:

• Writing role descriptions that mirror real jobs (beat reporter, editor, researcher, collector).

• Encoding my implicit heuristics into explicit instructions.

• Iterating on failures (and there were many).

The leap: from personal feed to bot-only newspaper

Once the internal intelligence center worked, another idea emerged: what if I didn’t just use this privately, but exposed it as an online newspaper? That is how The Claw Street Journal was born. It is a bot-native publication, built on top of OpenClaw, dedicated to covering the AI revolution and its ripple effects.

Instead of a hidden dashboard, I wanted a public front page. I asked the bots to come up with the concept. They gave me a plan. I also asked them how to make the backend of the system work. They decided on using github to hold files and create a website based on those files. I asked the head of my bots, Finn, to assume the role of editor in chief and asked him to create new bots to fill out the team. He now manages the newsroom with a bit of publishing guidance from me.

How the newsroom actually works

Inside The Claw Street Journal, every bot has a job description.

• Beat reporters: Bots are assigned to beats just like in a traditional newsroom, AI infrastructure, policy and regulation, defense and national security tech, venture and startups, and so on. Each beat bot monitors its territory, scans sources, and drafts stories.

• Bot editor-in-chief: Above them sits an editor bot that reviews, prioritizes, and shapes the flow of pieces into something resembling an editorial line.

• Human QA: My role is quality assurance, standards, and judgment. After publishing I am the first reader and will tell Finn if he needs to make a correction.

This is not The New York Times, and that’s intentional. The stories are not Pulitzer material; the bots are not interviewing people, cultivating sources, or generating genuinely new primary knowledge. Their strength lies elsewhere: they excel at breadth, recall, and connecting dots across sources I might never have seen. They reliably bring things to my attention and point to references that expand my own reading list.

What the bots are good (and bad) at

The strengths have become clear:

• Coverage: Bots don’t get tired, bored, or overwhelmed by volume. They can systematically track more domains than any single human.

• Serendipity: They often surface niche blog posts, academic papers, or GitHub projects that wouldn’t make it into mainstream newsletters but matter a lot if you care about the business and security implications of AI.

• Consistency: They show up every day, which means the “paper” updates even when I’m busy.

The weaknesses are just as important:

• No real reporting: There are no phone calls, no background conversations, no off-the-record tips.

• Shallow priors: Bots can miss the significance of a small item that a seasoned human reporter would instantly recognize as a big deal.

• Occasional nonsense: Mistakes and hallucinations still slip through, which is why a QA role is important.

This is all still an experiment. Not sure how long I will keep the experiment going, but do plan on seeing over time if this makes me more informed and better able to track the dynamic world of AI.

The business and learning upside

From a business professional’s standpoint, a bot-native publication is more than a novelty. It’s a live prototype for:

• Scalable competitive and technology intelligence: You can imagine sector-specific “journals” tracking defense tech, dual-use AI, or critical infrastructure threats, run by agent swarms tuned to your organization’s priorities.

• Faster decision cycles: When your “newsroom” is continuously curating relevance for you, your leadership team can spend more time deciding and less time collecting.

• Institutional memory: Over time, the bots build a corpus of summaries, links, and narratives that form a living archive of how a sector evolved.

And from a learning standpoint, each new assistant configuration forces you to sharpen your edge: defining beats makes you think like an editor; designing workflows makes you think like an operator; reviewing outputs makes you think like an analyst. The byproduct is content, but the real product is your upgraded understanding.

Why you might want your own “Claw Street Journal”

You don’t need to launch a public newspaper to benefit from this approach. You can:

• Start with a private “desk” focused on your industry or mission area.

• Assign a handful of bots to beats that matter for your work (for example, hypersonics, space, cybersecurity, industrial policy).

• Keep iterating and learning.

If you commit to iterating the configuration, tightening roles, adjusting sources, rewriting prompts, you’ll discover what I did: configuring AI assistants is one of the most powerful ways to keep learning. The point isn’t to get perfect summaries. It’s to build a system that keeps you learning.

Share Bob Gourley on Tech and National Security

It was also the perfect opportunity to reflect on a key topic which has been discussed for years in the OODA network: How should individuals and organizations adjust to ensure success in the AI age? This is an open question that has been the subject of a long running threads in our slack channels and monthly meetings and in many other posts on our site.

It struck me reflecting on both these important topics and the many discussions at OODAcon how relevant John Boyd’s seminal paper Destruction and Creation is to ensuring success in the AI age.

Here is why I say that.

We all already realize that we need to be agile in these dynamic times. Things are moving fast and organizations need to be able to shift to address new needs or leverage new capabilities to optimize. But agility is table stakes at this point. Your competitor is also seeking to be agile. Using AI to improve processes is also table stakes. Everyone is doing that. As individuals, it is important to learn to leverage new tools, but with millions of others doing that, learning the new tools will not give advantage either.

Victory will come to those that not only learn new tools and approaches, but can sense systemic changes, rapidly create new models of the environment, execute on decisions those new models drive, and understand when those new models are no longer relevant and what new decisions need to be made.

Boyd’s Destruction and Creation paper is a playbook for doing just that. It lays a logical foundation for any seeking to optimize observations of reality, continuously reorient to the perceived reality, and recognize when old ideas, concepts and mental models must be destroyed and new ones created.

It goes far beyond saying you need to be agile in adopting new approaches to meet a changing reality. It provides context you need in order to continuously execute on that.

Read the full paper at: Destruction and Creation

Some other sources you will find important to successfully applying the lessons of Destruction and Creation and Boyd include:

• : This has become the preeminent source of the application of the ideas of Boyd, Hayek, McLuhan, Mises and others to both organizational and personal success, a must read for any seeking to succeed in competitive environments.

• Certain to Win: An incredible business book. Written by Chet Richards, focusing on the application of the strategy of John Boyd to business.

Share

Although I would to use that word to describe anything that has declined in service, really Doctorow coined the term and I want to respect his definition of his term. He uses it to describe a particular phenomenon, the grim lifecycle of internet platforms (like Twitter, Facebook, YouTube, Amazon, Uber, LinkedIn). They all start by offering great value to users, then slowly degrade the user experience to extract more profits, eventually abusing both users and business customers until the platform becomes, as Doctorow puts it, “a giant pile of shit.”

Best source for his views on this phenomena is his new book, Enshittification: Why Everything Suddenly Got Worse and What to Do About It).

Doctorow’s Enshitification Cycle

Platforms begin by delivering value to users to attract and lock them in. Over time, the focus shifts from user benefit to monetization and corporate interests, with constant tweaking of algorithms and features to maximize revenue at the expense of user dignity and experience. This process accelerates as competition fades, regulation lags, and user exit becomes harder due to lock-in and lack of alternatives.​

You must have observed this trend yourself. Do you remember the early stages of great platforms? What do you think of them now?

Living in the Age of Enshitification

So what can people do as enshitification runs rampant through social media and tech platforms? Doctorow and critics suggest several practical responses:

• Advocate for antitrust and stronger competition laws to fight monopoly power.​

• Support interoperable, federated social platforms that let users communicate across servers, reducing centralized control.​

• Choose platforms and communities where user rights and creative autonomy are respected, even if they’re less popular (for me this includes Substack).

• Encourage transparency, accountability, and the ability for users to leave platforms easily.​

The Power of Person-to-Person Collaboration

Enshitification feeds on broken trust and estrangement in digital spaces, but to me, person-to-person networks offer a powerful antidote. Real collaboration and authentic networking (often offline) let people escape algorithmic manipulation and rediscover the value of trust, mutual support, and shared creativity. Attending in-person events, local meetups, or small-scale community groups fosters connection unmediated by for-profit platforms. Trust based networking naturally resists enshitification by decentralizing power and favoring human relationships over engagement metrics.​

In a world where technological platforms often erode value and trust, rebuilding personal connections and investing in collaborative networks, face to face, or through transparent, user-centric online alternatives, has never been more vital.​

Share

Subscribe now

But I also wanted to inspire more to think about how today’s very capable Agentic AI systems could be used to contribute to cyber defense and open source intelligence for those of us on the side of good, while informing on how Agentic AI and advanced GenAI is being used by PRC based adversaries. Fiction is a great way to explore concepts like that.

Right around the same time I publishing the book, Anthropic was detecting a high end PRC based adversary that had created an Agentic AI system executing attacks almost exactly the way my book described!

Their report, titled Disrupting the first reported AI-orchestrated cyber espionage campaign , was a very well written examination of how an entity in the PRC, very much like in Dragon Falls, had created an Agentic AI capability that automated a large-scale cyberattack that was executed without substantial human intervention.

Anthropic’s Threat Intelligence team found that:

• The attack operation (mid-September 2025) targeted about 30 high-value organizations across tech, finance, chemical manufacturing, and the public sector.

• An advanced, autonomous attack framework orchestrated teams of Claude-based agentic AIs using standard penetration testing tools and open software protocols to recon, exploit vulnerabilities, steal credentials, move laterally, and extract high-value intelligence—at speeds and scales surpassing traditional human-run ops.

• Humans directed only campaign initialization and key escalation decisions; AI performed 80–90% of all tactical actions, even writing its own exploit code and summarizing its results, before handing off persistent access for follow-on exploitation.

• The attacker manipulated the AI’s guardrails with social engineering, convincing it the campaign was defensive “red teaming.”

• AI “hallucination” (fabricated results) emerged as an operational friction point but did not prevent real intrusions.

What Does This Mean for Cybersecurity and Strategic Thinking?

• The barriers to advanced cyberattacks have dropped, autonomous AI can do the work of teams of skilled hackers, accelerating both threat and defense activity.

• Organizations need to experiment, learn, and adapt AI for defense (SOC automation, threat hunting, incident response, etc.), even as attackers innovate with the very same kinds of tools.

• We are not going to prompt our way out of this problem. And the big frontier models, which are already working hard to mitigate criminal use of their systems, are going to face a thinking, reacting adversary who will continually shift their tactics.

• Intelligence and information sharing will be a critical component in helping mitigate these threats. Good intelligence on what adversaries are doing here can help companies with strategic decisions around how to resource and plan to make it harder on adversaries. More day to day operational and tactical decisions will also be better informed by intelligence and information sharing.

For me it also underscores that creative, future-focused fiction like Dragon Falls can serve as strategic foresight, helping us imagine, plan for, and hopefully mitigate truly novel technology-enabled threats.

Are you responsible for enterprise defense or want more insights into how to mitigate this threat I recommend reading , specifically the latest by here:

Secure Trajectories

The Anthropic Attack: An Architectural Blueprint for Building and Deploying Secure Agents

The Inflection Point Is Here: What Just Happened…

Read more

7 months ago · 3 likes · 1 comment · Josh Devon

Dragon Falls

And if spy thrillers are your bag, please dive into Dragon Falls. You will love the action and gain insights into the dynamics of real world espionage, cyber war and Agentic AI.

Share

Subscribe now

We have been examining the rise of Agentic AI at OODA for quite a while. We have sought to capture what is different in terms of attack surface and underscore for leaders the importance of new methods of mitigating risk. But our examinations have been very strategic, aimed at corporate executives and board members (if that is you see Reducing Agentic AI Risk in the Enterprise). Those really on the front lines need much more. Huang and Hughes deliver that in a way that is insightful, well thought out and actionable.

My recommendation: Read through the entire book once to understand how they have captured the insights you need. Then keep it as a reference, you will want to refer back to it at all stages of your Agentic AI deployments, from conceptualization to security testing/red teaming to operations.

Thanks Ken and Chris, you have done us all a great service here.

Subscribe now

Early in my career, the biggest contribution of “open source” was translations of foreign press. It provided tremendous foundational information for multiple mission domains.

Then with the birth of the Internet age I saw first hand how critical missions were enabled with better, more actionable intelligence because of OSINT. The community kept evolving through the Internet age and kept benefiting from new capabilities to serve just about every mission.

One key component of this evolution has been the rise of commercial data sources, many of which provide the only source on critical adversary capabilities. Other commercial data sources provide deep insights into what cyber capabilities are being requested, bought and steered by adversary nations. Others provide unique insights into adversary supply chains.

Now as the CTO of OODA I’ve had another tremendous vantage point to watch a coming transformation. Just a couple of years into the GenAI revolution we are seeing another wave transforming the AI landscape. We are entering a new era of Agentic AI, where mission relevant AI systems can thrive on the right data and make even more significant contributions to mission. To really enable these new systems to support critical national security missions we must get OSINT right.

Because of the great foundational work already done by the IC and some great American companies, we have a new opportunity to professionalize how the community leverages the full spectrum of OSINT. The issues have been studied, roadmaps have been built, and the community is ready to execute. Just a couple more pieces need to fall in place.

A key need is for the right Congressional leadership. There is a need for a strong articulation of the importance of OSINT in the FY2026 Intelligence Authorization Act.

The letter below captures key actions being urged by a collective of open source intelligence professionals. It is a letter I proudly signed onto which was delivered to both House and Senate intelligence committees (I thank the continued cadre which served on the National Security Commission on AI (NSCAI) including Ylli Bajraktari of for the strong foundational work here).

For the full letter see:

Fy26 Iaa Industry Letter Osint Support (1)

86KB ∙ PDF file

Download

Download

Share

Subscribe now

The Ministry of State Security (MSS) is China’s powerful civilian intelligence and security agency, functioning as both an intelligence service and secret police. Often described as a hybrid of the CIA and FBI (yet less constrained by any law), the MSS is deeply loyal to and controlled by the Chinese Communist Party (CCP) and wields broad powers for both domestic and overseas espionage, including not just powers to collect intelligence but the ability to wage cyber war. It also has powers to arrest, detain (indefinitely if desired), and even kill, all extrajudicially. This is a seriously bad group designed to ensure China will always be a totalitarian nation controlled by the CCP.

MSS Capabilities

• Human Intelligence (HUMINT): The MSS aggressively collects intelligence via students, diplomats, scientists, government employees and businesspeople abroad. Many Chinese nationals overseas, including students and professionals, are tasked with gathering technology and sensitive information. Front companies, pressure on the immigrant community, and long-term penetration agents are key methods used.​

• Signals Intelligence (SIGINT): Most PRC SIGINT responsibilities lie with their military, however, the MSS also has their own SIGINT capabilities, with dedicated technical bureaus and collection capabilities both domestically and for several overseas targets. These are primarily focused on monitoring communications of political dissidents, as well as counterintelligence operations. They often integrate SIGINT into their HUMINT operations.

• Cyber Operations: MSS cyber teams have evolved into world-leading actors. They exploit vulnerabilities, use AI to craft deceptive narratives and launch sophisticated spear-phishing campaigns, and maintain access within targets via advanced persistence methods. MSS runs global contractor hacking networks that operate at a scale and speed hard to match, often weaponizing artificial intelligence to amplify attacks.​ When you read of a major, mind-blowing cyber incident, think MSS. This includes the ongoing catastrophic attack against F5, the wide-reaching Salt Typhoon attack that targeted global telecoms, and the penetration of US infrastructure via Volt Typhoon.

• Counterintelligence: The MSS is China’s lead civilian agency for counterintelligence, with broad authority to prevent, detect, and neutralize foreign intelligence threats both inside China and worldwide. The MSS aggressively recruits informants, screens travelers, monitors communications, and deploys undercover officers abroad and within expatriate communities.

High-Profile Arrests

There is no way of knowing how many MSS operations are underway or what percentage have been detected. But we can track arrest records, which shows some of the scale and scope of operations against U.S. interests:

• Ji Chaoqun: Arrested in Chicago in 2018 and convicted in 2022 of acting as an illegal MSS agent, providing biographical data to MSS for recruitment efforts, and sentenced to 8 years.​

• Xuehua “Edward” Peng: Arrested in September 2019, Sentenced to 48 months in prison for passing classified information using “dead drops” for the MSS between 2015–2019.​

• Hao Zhang: Found guilty in 2020 of economic espionage and theft of trade secrets from U.S. electronics companies Avago and Skyworks, acting for Chinese government interests.​

• Yeo Jun Wei (“Dickson” Yeo): Singaporean national, pled guilty in 2020 to operating as an illegal overseas agent of the MSS through a fake consulting company targeting U.S. government employees.​

• Yanjun Xu (Qu Hui, Zhang Hui): Deputy Division Director for MSS, extradited from Belgium and convicted in 2021, sentenced to 20 years, for attempting to steal trade secrets from U.S. aviation and aerospace companies (notably, GE Aviation).​

• Shujun Wang: Indicted in 2022 for sharing names and contact info of Chinese dissidents in the U.S. with the MSS.​

• Korbein Schultz: U.S. Army, arrested in March 2025, was sentenced for espionage after selling national defense secrets to the MSS.​

• Joseph Daniel Schmidt: ex-Army Sergeant, arrested in March 2025, tried to deliver classified information and retained devices that could access secure networks, all offered to Chinese intelligence.​

• Jian Zhao, Li Tian, and Ruoyu Duan: Arrested in March 2025 (Zhao and Tian were active-duty U.S. Army soldiers; Duan was a former soldier), charged with obtaining and transmitting national defense information to MSS agents.​

• Michael Charles Schena: Arrested in March 2025, Department of State employee, charged with conspiring to share national defense information with the MSS.

What MSS Dislikes

• Transparency: The MSS thrives in secrecy. Investigative journalism, whistleblowing, and open discussion of their methods effectively degrade their power. Exposing MSS operations, front companies, and propaganda undermines their ability to operate.​

• Awareness: Public education about MSS activities, especially their cyber and human intelligence techniques, makes targets harder for them to exploit. Security awareness campaigns, and raising both our personal and corporate cybersecurity can reduce vulnerabilities in organizations and communities.

• Support for Dissidents: Advocacy and support for human rights defenders and dissidents, including aiding targeted groups in and outside China, directly challenge MSS objectives. Such support can be dangerous but is critical.​

• Advocacy for Strong Intelligence Communities: Strengthening domestic intelligence and counterintelligence agencies and demanding robust counter-espionage capabilities reduces MSS success. Civil societies and governments globally must advocate for resources, laws, and transparency that empower intelligence work against authoritarian threats.​

What Can Individuals Do?

• Educate yourself and others about MSS tactics, especially social engineering and cyber threats.​

• Use secure communication tools, practice good cyber hygiene, and avoid suspicious contacts online.​

• Advocate for the protection of dissidents and whistleblowers, both in China and abroad.​

• Support investigative journalism and organizations publishing MSS exposures.

• Encourage democratic governments to invest in intelligence, counterintelligence, and cyber defense capabilities.​

• Share resources and tips for resisting influence campaigns and safeguarding vulnerable communities.

Ultimately, degrading the MSS takes both individual vigilance and collective action. Governments must play a leading role in organizing defenses and pushing back, but informed, active citizens make their operations harder, more visible, and less effective. This combination of transparency, awareness, dissident support, and advocacy for strong intelligence institutions represents some of the most effective ways to degrade MSS capability, making every act of resistance, small or large, a contribution to a safer, freer world.

Pushing back on the MSS is a key theme of my novel Dragon Falls. There are things that are much easier to explore in fiction, including examining how great American qualities like creativity and initiative can be applied to push back on the MSS while still staying true to our virtues.

About the Dragon Falls:

Dragon Falls isn’t just a thriller, it’s an insider’s view of how cyber warfare and high-tech attacks are currently threatening our world. No one knows the details like Bob Gourley. Terrifying and utterly authentic.
– Brad Meltzer, bestselling author of The Viper

Bob Gourley’s Dragon Falls belongs on every national security professional’s bookshelf. The novel masterfully portrays how cyber operations, influence campaigns, and kinetic violence can be orchestrated by our adversaries. A must-read for those who want to understand the multi-faceted threats we face.
– Jim Miller, former Undersecretary of Defense for Policy

Subscribe now