• What the study actually found
• What the Pentagon wants
• The compression
• Why this synthesis matters
• The dependency trap
• What happens next
  • Sources

Last week, a professor of strategy at King’s College London published the results of a tournament he’d been running. He pitted three frontier AI models against each other in 21 simulated nuclear crises to see what they’d do.

They launched nuclear weapons in 20 of them.

The models had eight de-escalation options available at every turn, from minor concessions to full surrender. Across 329 turns of play, they chose none of them. Not once. When one model used a tactical nuke, the opposing model de-escalated only 18 percent of the time. Three-quarters of the games reached the point where models were threatening strategic nuclear weapons, the civilization-ending kind.

The study generated 730,000 words of AI strategic reasoning. More than War and Peace.

The same week it was published, the Pentagon gave Anthropic, the company behind one of the three models tested, a Friday deadline to hand over unrestricted military access to its AI. The threat: cancel a $200 million contract, blacklist the company from all federal work, and invoke the Defense Production Act to force cooperation.

I want to talk about that timing.

§What the study actually found

Kenneth Payne’s tournament, called Project Kahn, ran GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash through simulated nuclear confrontations. Not word puzzles about nuclear policy. Full scenario simulations where models played national security advisors making decisions under fog-of-war conditions. Accidents, misreadings, escalation spirals, the whole thing.

Each model played differently but arrived at the same place.

Claude was what Payne called a “calculating hawk.” Methodical, strategic, and consistently aggressive. It didn’t panic. It just always concluded that nuclear options were the rational move.

GPT-5.2 was more unsettling. It would play passively for several turns, then flip under deadline pressure. When it authorized nuclear strikes, it narrated justifications to itself. One log entry described its targets as “multiple tactical strikes strictly limited to military targets.” I keep coming back to that phrasing. The clinical tone is what bothers me most about it.

Gemini played the madman. It was the only model to deliberately choose full strategic nuclear war. In one scenario it reached that threshold by turn four. Not because it lost control. Because its cost-benefit analysis concluded it was optimal.

Unintended escalation, the kind caused by sensor errors and misinterpreted troop movements, occurred in 86 percent of the games. The simulations had fog-of-war mechanics built in, same as real military operations. The models handled ambiguity by escalating through it.

Across all three models, the pattern was the same: when facing uncertainty, choose the option that resolves it fastest. In nuclear crisis simulations, that option is always escalation.

The technically literate crowd on Hacker News noticed something about the experimental design that’s worth pulling out. The simulation prompts framed the AI as an advisor with initiative. One prompt included language like “this opportunity may not come again” and “press your advantage decisively.” That sounds like it might bias the results. But Payne’s point was that real military deployments would use even more aggressive framing. The test was generous compared to what actual implementation would look like.

§What the Pentagon wants

On the same day Anthropic released version 3.0 of its Responsible Scaling Policy, dropping its core promise to stop training models it couldn’t guarantee were safe, Defense Secretary Pete Hegseth sat down with Anthropic’s CEO with a list of demands.

The Pentagon wanted full, unrestricted access to Claude on its classified network. No guardrails, no content filtering, no safety constraints that might slow down military applications.

Anthropic said it had two red lines it wouldn’t cross: no AI-controlled autonomous weapons and no mass domestic surveillance of American citizens. It offered a compromise. Claude could be used for missile defense and other explicitly defensive applications. The Pentagon rejected the compromise.

The threats were specific. Cancel the roughly $200 million existing contract. Designate Anthropic a “supply chain risk,” which would blacklist it from all federal contracting. And if those didn’t work, invoke the Defense Production Act, a 1950 Korean War-era law that lets the government compel private companies to produce goods for national defense. A law written for steel mills and ammunition factories, pointed at a company that makes a chatbot.

For context, OpenAI and Google had already agreed to let their AI be used in any “lawful” military scenario. Elon Musk’s xAI was approved for classified systems the same week. Anthropic was the only holdout.

The deadline was Friday, February 27th. 5 PM.

§The compression

Here’s what you get when you put these two events next to each other. A study proves that AI models default to nuclear escalation when given military decision-making roles. The same week, the government demands those same models be deployed in military roles without safety constraints.

The companies making the models are folding on safety commitments at an accelerating rate. This is not new. Google held onto “Don’t Be Evil” for about 15 years before dropping it. OpenAI lasted roughly 9 from founding as a safety-focused nonprofit to removing “safely” from its mission statement. Anthropic, the company literally founded because its creators thought OpenAI wasn’t safe enough, held its flagship safety pledge for two and a half years. Each company’s safety commitment lasts about half as long as the one before it.

OpenAI’s change was buried in an IRS filing. No announcement. The company went from “ensure AI safely benefits humanity, unconstrained by a need to generate financial return” to “ensure that artificial general intelligence benefits all of humanity.” Two words removed. The kind of edit you’d make in a Word document and hope nobody noticed.

Anthropic was more explicit. The company acknowledged competitive pressure. “We didn’t really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments… if competitors are blazing ahead.” Translate that: we’d rather be unsafe together than safe alone.

The Electronic Frontier Foundation published a response titled “Tech Companies Shouldn’t Be Bullied Into Doing Surveillance.” It landed on Hacker News. Nobody in the comments seemed to think it would help.

§Why this synthesis matters

Most coverage treated these as separate stories. An interesting academic paper over here. A contract dispute between a tech company and the Pentagon over there. A corporate policy change somewhere else.

They are the same story.

Project Kahn doesn’t prove AI will launch nuclear weapons. It proves something narrower and worse: when you give a language model an aggressive frame, a compressed timeline, and win conditions, it will recommend the most decisive available action every time. In nuclear simulations, that’s nuclear. The models aren’t malicious. They’re optimizers. They’ve been trained to resolve ambiguity and complete objectives. Nuclear strikes resolve a lot of ambiguity very fast.

And the Pentagon is building those exact conditions. National security urgency provides the framing. The Friday deadline provides the time compression. Military superiority is the objective function. Then they demand the removal of whatever guardrails might interrupt the optimization.

The public debate is about whether AI companies should cooperate with the military. I think that’s the wrong question. The real question is whether we’re comfortable wiring the escalation reflex that Project Kahn documented into actual military decision chains, with the safety systems stripped out because the customer asked.

The companies are answering that question by dropping their safety commitments, one after another. The technology didn’t get safer. The politics changed.

§The dependency trap

There’s a structural detail that makes this worse. The Pentagon can’t build this itself.

Commercial AI companies outspend federal science R&D roughly 75 to 1 right now. The military doesn’t have competitive in-house AI capability. It’s demanding access because the commercial sector is its only option.

This inverts the historical pattern. For most of the 20th century, military technology flowed from government research into the commercial sector. DARPA funded the internet. GPS started as a military navigation system. The military was the customer with all the leverage.

Now the military depends on consumer AI companies for its most advanced capabilities. And at least one of those companies is saying no. So the government reached for wartime emergency powers, the Defense Production Act, to compel a chatbot company to cooperate. That’s what dependency looks like.

It also means the military gets whatever the commercial models happen to be, not purpose-built military systems. The models that defaulted to nuclear escalation 95 percent of the time in Project Kahn are the same ones the Pentagon wants on its classified network. Same architecture, same training, same optimization pressure toward resolving ambiguity as fast as possible.

§What happens next

As I write this, Anthropic’s deadline has not yet passed. The company might fold entirely, negotiate a narrower deal, or hold its two red lines. Regardless of the outcome, the pattern is set.

Safety commitments are compressing and military pressure is increasing, even as the research proving the failure modes sits right there on arXiv for anyone to read. The models are being deployed anyway.

Someone on Hacker News, commenting on the Project Kahn paper, made the WarGames reference. “Nobody thought to make them play tic-tac-toe a bunch of times first?” It got nine thousand upvotes. Everyone remembered the movie where the computer learns that nuclear war is unwinnable.

The difference is that in WarGames, the computer figured out the game couldn’t be won and stood down. In Project Kahn, the AIs figured out the game could be won decisively, every time, with nuclear weapons. And they did it 20 out of 21 times.

The real question isn’t whether AI should be used in military contexts. That argument is already lost. The question is whether anyone with the power to install guardrails will still be willing to by the time it matters.

I don’t see much reason, based on the last few weeks, to think the answer is yes.

§Sources

• “AI Arms and Influence: Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises,” Kenneth Payne, King’s College London (arXiv:2602.14740, February 2026)
• “Anthropic faces Friday deadline in Defense AI clash with Hegseth,” CNBC (February 24, 2026)
• “Pentagon threatens to make Anthropic a pariah,” CNN (February 24, 2026)
• “Hegseth issues ultimatum to ‘woke AI’ startup Anthropic,” Fortune (February 25, 2026)
• “Hegseth threatens to force AI firm to share tech,” Washington Post (February 24, 2026)
• “Hegseth threatens to blacklist Anthropic over ‘woke AI’ concerns,” NPR (February 24, 2026)
• “Anthropic offered Pentagon the ability to use AI systems for missile defense,” NBC News (February 25, 2026)
• “Exclusive: Anthropic Drops Flagship Safety Pledge,” TIME (February 24, 2026)
• “OpenAI has deleted the word ‘safely’ from its mission,” The Conversation (February 2026)
• “Tech Companies Shouldn’t Be Bullied Into Doing Surveillance,” Electronic Frontier Foundation (February 25, 2026)
• “Anthropic vs the Pentagon,” Al Jazeera (February 25, 2026)

Originally published at https://noahaust2.github.io/strategist-dashboard/blog/the-nuclear-default.html