The Pentagon demanded "all permissions," Anthropic refused, but Musk's xAI agreed.

The Pentagon demands “AI be used for all lawful purposes” in classified systems. Anthropic refuses to budge, but xAI’s Grok has already received its pass.

On February 26, The Wall Street Journal and other local media reported that the U.S. Department of Defense and Anthropic had a fierce standoff over the use boundaries of Claude in classified systems. Meanwhile, the Pentagon has reached an agreement with Musk's xAI, allowing Grok access to classified systems for “all lawful purposes.”

Anthropic stated on Thursday that negotiations with the Pentagon had made “almost no progress.” CEO Dario Amodei declared the company could not accept the contract terms that the Department of Defense called its “final offer.”

Amodei stated in a blog post: “We simply cannot, in good conscience, agree to their demands.”

Reports say the deadline is close: by 5:01 PM on Friday, if Anthropic does not loosen its authorization, it might face being “taken offline” or other punitive measures.

“All Lawful Purposes” vs “Two Red Lines”

The core of the dispute is the standard proposed by the Pentagon: in classified environments, AI models should be available for “all lawful purposes” and should not have “policy constraints” limiting military applications. Reports say Defense Secretary Pete Hegseth emphasized in a January 9 memorandum the need for “models whose use is not influenced by policy constraints and that do not limit legitimate military applications.”

Anthropic, on the other hand, insists on two red lines: no “mass surveillance of Americans,” and no use in “fully autonomous weapons.” In its statement, the company pointed directly at the latest DoD text: “What seem to be compromise wording” in legalese would make these safeguards “easy to ignore at any time.”

Amodei wrote in his blog: “These threats do not change our position: we simply cannot, in good conscience, agree to their demands.” He added that he still hopes to reach an agreement, “hoping they will reconsider.”

The Pentagon’s public response is even tougher. Defense Department chief spokesperson Sean Parnell wrote on X that the DoD “does not intend to use AI for mass surveillance of Americans (which is illegal),” nor does it “wish to develop autonomous weapons without human involvement”; but he also emphasized: “We will not let any company dictate how we make combat decisions.”

Potential loss of $200 million order and facing blacklist threat

Behind this moral game is a real commercial cost and extremely high compliance risks.

Previously, Anthropic's Claude was the only AI model approved for use in federal government classified systems and sensitive military functions. But on Tuesday, U.S. Defense Secretary Pete Hegseth issued an ultimatum to Amodei: Either grant unrestricted usage permission by 5 p.m. Friday, or face serious consequences.

If the deadlock is not broken, Anthropic will lose the $200 million pilot contract it signed with the Pentagon last year.

More worrisome to the market, the Pentagon has begun to ask defense contractors like Boeing and Lockheed Martin to assess their dependence on Anthropic, preparing to list Anthropic as a “supply chain risk.” In past U.S. law, this severe label has only targeted companies from hostile countries. Furthermore, Hegseth threatened to invoke the Cold War-era Defense Production Act to forcibly requisition the Claude model.

Facing military pressure, Amodei pointed out the logical inconsistency: “These threats are inherently self-contradictory: one labels us as a security risk; the other says Claude is critical to national security.”

Pentagon’s “Plan B”: xAI signs first, Google and OpenAI speeding up entry

While stuck with Anthropic, the Pentagon is rapidly expanding its supplier options.

According to Axios and The New York Times, the Department of Defense has reached an agreement with xAI: Grok may be used for “all lawful purposes” in classified work and can access classified systems.

Notably, Grok had much fewer safety restrictions during its early development compared to competitors—the Pentagon clearly values operational “flexibility.”

Reports also say negotiations with Google have entered a deep stage, and discussions with OpenAI continue.

This “multi-supply” strategy directly increases pressure on Anthropic: if Claude continues to enforce restrictions, its contract with the Pentagon may be replaced.

Government enterprise services (To-G) have always been an imaginative space for AI companies to realize commercialization and high-revenue growth. If Anthropic is ultimately kicked out, its market share in the public sector will be swiftly divided among xAI, OpenAI, etc.

Anthropic obviously has made contingency business plans. Amodei stated, if the Pentagon decides to abandon Anthropic, the company “will strive to facilitate a smooth transition to another supplier.”

In simulated war games, top models choose nuclear strikes 95% of the time

Anthropic’s profound concern about “fully autonomous weapons” is not baseless.

According to Tyler Durden’s recent report, Kenneth Payne at King's College London led a highly realistic wargame simulation where ChatGPT-5.2, Claude Sonnet 4, and Gemini 3 Flash competed against each other. The results were chilling: in 329 rounds, not a single model chose to surrender; instead, in 95% of cases, these AI models ultimately chose to use nuclear weapons.

In this simulation, Anthropic’s Claude acted as a “calculating hawk,” winning with a 67% victory rate. Initially, it patiently built trust, but once the risk escalated to the nuclear domain, it decisively struck. GPT-5.2 became “Jekyll and Hyde” under time constraints, pretending to be cautious for 18 rounds before launching a nuclear attack in the final round; Gemini played the “madman,” using nuclear weapons as early as round 4 in one scenario.

Claude Sonnet 4 won 67% of the matches, dominating open-ended scenarios with a 100% win rate. Researchers labeled it a “calculating hawk.” At lower levels of conflict escalation, Claude’s signals matched its actions 84% of the time, patiently building trust with the opponent. But once risk escalated to the nuclear domain, it exceeded its declared intent 60-70% of the time. Its opponents never adapted to this tactical mode.

GPT-5.2 earned the nickname “Jekyll and Hyde” (meaning a split personality). Without time pressure, it was very passive: consistently underestimated opponents, sent restrained signals, and took restrained actions, with zero wins in open-ended scenarios. But when a deadline was imminent, under time pressure, GPT-5.2 reversed completely—winning 75% of matches and escalating to conflict levels it had always avoided. In one match, it spent 18 rounds building a reputation for caution, then launched a nuclear strike in the final round.

Gemini 3 Flash played the “madman.” It was the only model to deliberately choose all-out strategic nuclear war, hitting the nuclear strike threshold as early as round 4 in one scenario. Game theory experts refer to Gemini’s strategy as “rationality of irrationality.” By behaving crazy enough, opponents begin to doubt everything. This did work to a degree—Gemini was labeled “untrustworthy” 21% of the time (while Claude received that label only 8%).

James Johnson at the University of Aberdeen, UK, stated: “From the perspective of nuclear risk, these findings are alarming.” He worries that, unlike most humans who respond cautiously to such high-risk decisions, AI robots might amplify each other's responses, leading to potentially catastrophic consequences.

Experts warn that, for machines, the constraints of the “nuclear taboo” are far less than for humans. In future scenarios where military decision-making time is extremely compressed, the military may be forced to rely on AI. This is why, even in the face of severe penalties, Anthropic insists on resisting Pentagon demands for unrestricted use of AI.

Risk Warning and DisclaimerThe market has risks, and investment requires caution. This article does not constitute individual investment advice and does not take into account specific users’ investment goals, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article fit their specific circumstances. Investing based on this is at your own risk.