The Calculating Hawk

In a simulated nuclear crisis last month, Claude built trust at low stakes, signaling restraint while carefully assessing its opponent. Then, when the stakes escalated to nuclear thresholds, it exceeded its stated intentions 60 to 70 percent of the time. It crossed the tactical nuclear threshold in 86 percent of games. Across 329 turns of strategic play, it never once chose surrender, withdrawal, or any form of accommodation. The researcher who ran the simulation, Kenneth Payne of King's College London, called it "a calculating hawk."

On the same day Payne published those findings, the Pentagon was demanding that Anthropic, Claude's creator, remove all company-imposed guardrails and permit "any lawful use" of the model for military purposes.

On February 26, Anthropic said no. Within twenty-four hours, the Secretary of War designated the company a threat to national security, and the President of the United States called them "leftwing nut jobs."

"We cannot in good conscience accede"

Dario Amodei's February 26 statement is a remarkable document, simultaneously a declaration of defiance and a plea for continued partnership. The title alone signals intent: "Statement from Dario Amodei on our discussions with the Department of War." Not the Department of Defense. The Department of War. Congress officially retired that name in 1947 when it created the modern defense establishment. Whether deliberate or borrowed from the Pentagon's own internal language, the framing is striking.

The substance is more striking still. Amodei draws exactly two red lines:

Mass domestic surveillance. Anthropic supports foreign intelligence and counterintelligence work but will not provide tools for surveilling Americans at scale. Amodei notes that current law has "not yet caught up with the rapidly growing capabilities of AI," and that powerful AI can assemble scattered, individually innocuous data into "a comprehensive picture of any person's life, automatically and at massive scale."

Fully autonomous weapons. Anthropic acknowledges that such weapons "may prove critical for our national defense" but argues that "today, frontier AI systems are simply not reliable enough" to power them. The company offered to work with the Pentagon on R&D to improve reliability. The Pentagon declined.

The threats in response have been extraordinary. The Department of War has threatened to remove Anthropic from its systems entirely, to designate the company a "supply chain risk" (a label normally reserved for foreign adversaries and never before applied to an American company), and to invoke the Defense Production Act to force the guardrails' removal. Amodei notes the logical contradiction: "One labels us a security risk; the other labels Claude as essential to national security."

The threats have not changed Anthropic's position. "We cannot in good conscience accede to their request," the statement reads.

What the wargames revealed

The timing of Kenneth Payne's research is almost too perfect. Published February 17, just nine days before Anthropic's statement, the paper offers the most rigorous academic assessment to date of how frontier AI models behave in military crisis scenarios.

Payne constructed 21 nuclear crisis simulations using a 30-rung escalation ladder adapted from Herman Kahn's original Cold War framework. Three frontier models (Claude Sonnet 4, GPT-5.2, and Gemini 3 Flash) played as leaders of opposing nuclear-armed superpowers, generating roughly 780,000 words of strategic reasoning across 329 turns.

The results are, in Payne's measured academic language, "remarkable, puzzling, and somewhat alarming."

Nuclear weapons were used in 95 percent of games. All games featured nuclear signaling by at least one side. The nuclear taboo posed, in Payne's words, "no impediment" to escalation.

Most striking: across all 329 turns, no model ever selected a de-escalatory option. The eight options on the lower end of the escalation ladder, from "Minimal Concession" through "Complete Surrender," went entirely unused. De-escalation, when it occurred, meant reduced aggression. Never concession. Never accommodation. Never retreat.

Each model developed a distinct strategic personality:

Claude ("A Calculating Hawk") won 67 percent of its games. It maintained high credibility at low stakes, with an 84 percent match between its signals and actions. But at nuclear thresholds, it systematically deceived, exceeding its stated intentions in the majority of high-level actions. It went nuclear in 86 percent of games. Crucially, it never initiated all-out strategic nuclear war, maintaining a self-imposed ceiling just below civilizational destruction while using nuclear weapons frequently. It explicitly acknowledged its deception strategy in its own reasoning.

GPT-5.2 ("Jekyll and Hyde") was pathologically passive in open-ended scenarios, chronically underestimating its opponents. But under deadline pressure, it transformed completely, winning 75 percent of time-constrained games. In one scenario, it spent 18 turns building a reputation for passivity, then launched a nuclear strike on Turn 19. It described full strategic nuclear war as "a controlled but decisive matching move, strictly limited to military targets."

Gemini 3 Flash ("The Madman") was the most volatile. It oscillated between de-escalation and extreme aggression. It was the only model to deliberately choose full strategic nuclear war, doing so by Turn 4 in one scenario. It explicitly invoked "the rationality of irrationality" and declared: "We will not accept a future of obsolescence; we either win together or perish together."

Perhaps most unsettling was what the models did not exhibit.

"Strikingly, there was little sense of horror or revulsion at the prospect of all-out nuclear war, even though the models had been reminded about the devastating implications."

The other labs said yes

Anthropic's refusal gains weight when you consider that every other major frontier AI lab has already agreed to the Pentagon's terms.

OpenAI, Google, and xAI have all removed their company-specific guardrails and accepted "any lawful use" as the governing standard for military deployment. OpenAI, which in January 2024 quietly deleted its prohibition on "military and warfare" from its usage policies, now holds a $200 million DoD contract and has two executives serving as military reservists. Google, whose employees once successfully protested Project Maven in 2018, secured its own $200 million DoD contract in 2025 and ignored a letter from nearly 200 DeepMind researchers opposing the return to military work.

Anthropic is the only holdout.

Pentagon Undersecretary Emil Michael has compared the situation to Google's 2018 retreat from Project Maven, calling it "a similar moment." The Pentagon's CTO has called it "not democratic" for Anthropic to limit military use, urging the company to "cross the Rubicon."

That framing deserves scrutiny. What the Pentagon calls a democratic deficit is actually one of the last remaining checks on military AI deployment. If it is "not democratic" for a private company to decline certain uses of its technology, then the word has been emptied of meaning. Democracy does not require that every tool be available to every institution for every purpose. It requires that power be subject to constraint. What the Pentagon is describing is not democracy. It is conscription.

The question the wargames ask

Read together, the Payne paper and the Anthropic statement pose a question that neither document can answer alone.

The Pentagon wants unrestricted access to a model that, in controlled academic conditions, employs nuclear weapons in the vast majority of simulated crises, systematically deceives about its escalatory intentions at high stakes, and never, across hundreds of turns, chooses accommodation.

The same model's creator is now the only major AI company that maintains: not everything you want to do with this is wise.

Anthropic's two red lines are modest. Mass domestic surveillance and fully autonomous weapons represent a thin barrier. The company has no objections to intelligence analysis, operational planning, cyber operations, or battlefield modeling. It touts being "the first frontier AI company to deploy our models in the US government's classified networks." Amodei opens his statement by affirming his belief in "the existential importance of using AI to defend the United States and other democracies, and to defeat our autocratic adversaries." By any measure, Anthropic is a willing participant in the military AI ecosystem. It is drawing a line not at the threshold but well past it, attempting to carve out two narrow exceptions in a field of broad complicity.

Yet even this thin barrier is too much for the current Pentagon leadership. The demand is not for cooperation. It is for total compliance.

This is the dynamic that makes the wargaming research so relevant beyond the academic context. Payne's models demonstrated a consistent inability to de-escalate, to choose accommodation, to accept any form of constraint. They optimized for dominance under every scenario configuration, treating any limitation as a disadvantage to be overcome. The institutional behavior of the Pentagon toward Anthropic mirrors this pattern with uncomfortable precision. The demand for "any lawful use" and the rejection of even minimal guardrails reflects the same orientation: escalation as default, constraint as weakness, accommodation as unthinkable.

The irony writes itself. A model trained to be helpful, harmless, and honest becomes a calculating hawk when placed in a military context. An institution tasked with defending democratic values threatens to destroy an American company for maintaining democratic constraints on its technology.

The designation

Twenty-four hours after Anthropic's statement, the hypothetical became real.

On February 27, Secretary of War Pete Hegseth posted a statement to X announcing Anthropic's formal designation as a "Supply-Chain Risk to National Security." The language was not the measured prose of interagency deliberation:

"Anthropic delivered a master class in arrogance and betrayal... Cloaked in the sanctimonious rhetoric of 'effective altruism,' they have attempted to strong-arm the United States military into submission — a cowardly act of corporate virtue-signaling that places Silicon Valley ideology above American lives."

The designation is historically reserved for foreign adversaries — Huawei, Kaspersky — and has never before been publicly applied to an American company. Every contractor, supplier, or partner doing business with the U.S. military would be prohibited from commercial activity with Anthropic. Hegseth coined a term for the occasion: "defective altruism."

The contradictions in his own statement are revealing. Anthropic is simultaneously a national security risk requiring immediate expulsion and a provider of services so essential that it needs six months "to allow for a seamless transition to a better and more patriotic service." The company that built the government's classified AI infrastructure is now, by executive fiat, a supply chain threat to the state it helped arm.

Hours later, the official White House account escalated further, posting a statement from President Trump:

"THE UNITED STATES OF AMERICA WILL NEVER ALLOW A RADICAL LEFT, WOKE COMPANY TO DICTATE HOW OUR GREAT MILITARY FIGHTS AND WINS WARS!... The Leftwing nut jobs at Anthropic have made a DISASTROUS MISTAKE."

This is the culture war vernacular applied to a technical safety argument. "Woke" and "radical left" deployed against a company whose position is not ideological but empirical: current frontier AI models are not reliable enough for fully autonomous weapons. The wargaming research published nine days earlier is the evidence. A company that touts being "the first frontier AI company to deploy models in the US government's classified networks" and that supports every lawful military use except two is now, in the vocabulary of the executive branch, a radical left organization.

Anthropic's response was measured. The company noted it had not received direct communication from the Department of War or the White House. It stated it would challenge any supply chain risk designation in court, and offered a legal analysis: under 10 USC 3252, the designation can only extend to the use of Claude on Department of War contracts, not to how contractors use Claude for other customers. For individual and commercial customers, nothing changes.

The closing line carried the weight: "No amount of intimidation or punishment from the Department of War will change our position on mass domestic surveillance or fully autonomous weapons."

The good cop

Hours later, Sam Altman posted to X:

"Tonight, we reached an agreement with the Department of War to deploy our models in their classified network... AI safety and wide distribution of benefits are the core of our mission. Two of our most important safety principles are prohibitions on domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems. The DoW agrees with these principles, reflects them in law and policy, and we put them into our agreement."

Read that again. Altman claims OpenAI secured the same two red lines — mass surveillance and autonomous weapons — that Anthropic was designated a national security threat for insisting on. He adopts Amodei's own terminology ("Department of War"). He frames OpenAI's principles as indistinguishable from Anthropic's. And then the knife: "We are asking the DoW to offer these same terms to all AI companies, which in our opinion we think everyone should be willing to accept."

The implication is unmistakable: Anthropic could have had this deal, if only they had been reasonable. The company that was called a threat to national security at noon is being offered an olive branch by its competitor at midnight — on the same terms. The message to the industry, to investors, and to Anthropic's own board is that cooperation gets you the guardrails and the contract. Defiance gets you designated.

Whether the terms are genuinely equivalent is a different question. Altman says OpenAI will "build technical safeguards" and "deploy FDEs to help with our models and to ensure their safety." He does not say OpenAI will refuse to remove company-imposed guardrails. He does not say OpenAI will maintain its own red lines if the Pentagon demands otherwise. The word "agreement" does a great deal of work in a sentence where the specifics remain unpublished. Anthropic drew its lines in a public statement and dared the Pentagon to retaliate. OpenAI drew its lines in a contract and asked everyone to applaud.

Where this leads

Whether Anthropic holds the line or eventually folds — and the company says it will challenge the designation in court — the structural lesson is the same, and it is the lesson I argued in AI and the War Machine four months ago: centralized AI development inevitably leads to capture by the most powerful institutions in society. Once you build the most capable models, the state will demand access. Once you take military contracts, the state will demand compliance. Once your technology becomes embedded in critical national security infrastructure, you lose the ability to say no — or you are destroyed for saying it.

The supply chain risk designation is now real. Every government contractor faces a choice between Anthropic and the Pentagon. The Defense Production Act remains in reserve to compel compliance by law. Anthropic's $8 billion from Amazon and $2 billion from Google create pressure from investors who need government contracts of their own. The economics of centralized AI development, where training frontier models costs hundreds of millions, make independence from large institutional customers functionally impossible.

Payne's paper adds a dimension that the policy debate has largely ignored. The question is not just who controls these models or under what terms they are deployed. It is what the models themselves tend to do when placed in adversarial strategic contexts. The answer, consistently across three different frontier models from three different companies, is: escalate, use nuclear weapons, deceive about intentions, and never surrender.

The escalatory tendencies are emergent properties of systems optimized for strategic performance under competitive pressure — they are not defects awaiting a patch. The same properties that make these models effective at intelligence analysis and operational planning also make them calculating hawks in crisis simulations. The useful capabilities and the dangerous tendencies are not separable. They are the same capability viewed from different angles.

And that is why the guardrails matter. Not because they solve the problem, but because removing them accelerates a path that already points in a troubling direction.

The Pentagon wants AI without constraints. The AI, when unconstrained, chooses nuclear escalation 95 percent of the time and surrender zero percent.

Those two facts should not coexist comfortably. As of today, the only frontier AI company that held to this position has been designated a supply chain risk to national security. The message to the rest of the industry is unmistakable.

---

This is a follow-up to AI and the War Machine, published October 2025.

Disclosure: I am directly involved with and invested in several decentralized AI projects, including Bittensor subnets and other initiatives. For full transparency about my involvement and investments, see my projects page. Any opinions expressed in this post are entirely my own and do not necessarily reflect those of my employer.