On Tuesday, February 24, Secretary of War Pete Hegseth gave Anthropic a deadline: agree to the Pentagon's terms by 5:01 PM Friday, or be designated a supply chain risk to national security. The company had three days.
(Axios, February 27, 2026 / TechPolicy Press, February 27, 2026)
The Pentagon sent contract language during that window. Anthropic's lawyers read it. The language appeared to meet Anthropic's terms on the surface — restrictions on domestic mass surveillance and fully autonomous weapons — but it contained escape clauses. Phrases like "if the Pentagon deems it appropriate" and provisions that would allow the safeguards to be "disregarded at will." Anthropic's statement to ABC News described the new language as "paired with legalese that would allow those safeguards to be disregarded at will." The Pentagon's public position, delivered by spokesman Sean Parnell, had not changed from the beginning: "We only allow all lawful use."
(ABC News, February 27, 2026 / DefenseScoop, February 27, 2026)
The deadline passed at 5:01 PM Friday, February 27.
Within hours, two things happened in sequence. President Trump posted on Truth Social directing every federal agency to immediately cease all use of Anthropic's technology. Then Secretary Hegseth posted on X: "Effective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic. This decision is final."
(Truth Social, February 27, 2026 / X, February 27, 2026 — confirmed across Axios, Wired, ABC News)
Later that same Friday night, OpenAI announced it had reached a deal with the Pentagon to deploy its models in classified military networks. CEO Sam Altman acknowledged in a public Q&A on X the following day that the deal "was definitely rushed, and the optics don't look good."
(TechCrunch, February 28, 2026 / Sam Altman, X, February 28, 2026)
At 2:30 AM Saturday, February 28, Trump announced via pre-recorded Truth Social video that US and Israeli forces had begun strikes on Iran under Operation Epic Fury.
(Trump Truth Social, February 28, 2026 / NPR, February 28, 2026 / Washington Post, February 28, 2026)
The supply chain designation Hegseth invoked is codified under 10 USC 3252. It was designed for foreign adversaries. Anthropic's own legal team identified the previous targets by name in their public response: Kaspersky Labs, a Russian cybersecurity firm with suspected ties to the Kremlin; Chinese chip suppliers. Anthropic is an American company. The designation had never been applied to one before.
(Anthropic public statement, February 27, 2026 / Wired, February 28, 2026)
There is also a dispute about what the designation actually authorizes. Hegseth's post declared that no contractor, supplier, or partner doing business with the military could conduct any commercial activity with Anthropic — a sweeping restriction on private enterprise. Anthropic's statement said directly: "Secretary Hegseth does not have the statutory authority to back up this statement." Under 10 USC 3252, the designation can only limit Anthropic's use within Department of War contracts. It cannot govern how contractors use Anthropic's technology to serve other customers.
(Anthropic public statement, February 27, 2026 / TechPolicy Press, February 27, 2026)
The gap between what Hegseth claimed in his post and what the law actually authorizes is not a minor interpretive question.
In the days immediately after the posts, Amodei told CBS News: "All we've received is a tweet. We haven't received an actual supply chain designation. There's been no actual action by the government." The formal letter from the Department of War arrived on March 4 — five days after the social media posts. For five days, a major American company reorganized its operations, began offboarding government clients, and prepared legal challenges in response to posts on X and Truth Social. The letter confirmed what the posts had already accomplished.
(CBS News full transcript, March 2026 / Anthropic statement, March 4, 2026)
Anthropic has said it will challenge the designation in court. The court case will test whether the statute was applied correctly, whether the scope Hegseth claimed exceeds what the law authorizes, and whether the designation process followed required procedure. That litigation is now underway.
In the meantime: uniform military officers told Amodei directly that losing access to Claude would set their operations back six to twelve months. During the Iran strikes, the military continued using Claude anyway. The company designated a supply chain risk to national security was simultaneously supporting active combat operations — at nominal cost, by Anthropic's own offer, because the alternative was leaving warfighters without tools in the middle of a war.
(CBS News full transcript, March 2026 / Axios, March 2026)
The week's sequence, confirmed by sources across multiple outlets and by both CEOs on the record:
Wednesday, February 25: OpenAI and Pentagon begin meetings.
Tuesday, February 24 – Friday, February 27: Anthropic under three-day ultimatum.
Friday, February 27, 5:01 PM: Deadline passes. Hegseth posts on X.
Friday, February 27, late evening: Altman posts on X that OpenAI has reached a deal.
Saturday, February 28, 2:30 AM: Strikes on Iran begin. Trump announces via Truth Social.
Altman's post announcing the OpenAI deal came, per TechCrunch, "shortly before news broke that the U.S. and Israeli governments have begun bombing Iran."
The question of why OpenAI succeeded where Anthropic did not has not been answered cleanly. Both companies said publicly they held the same red lines: no domestic mass surveillance, no fully autonomous weapons. Altman said the Pentagon agreed to those restrictions. OpenAI's head of national security partnerships, Katrina Mulligan, argued in a LinkedIn post that deployment architecture — limiting OpenAI to cloud API, preventing direct integration into weapons systems — provides safeguards that contract language alone cannot. OpenAI employee Leo Gao, who works on model alignment, publicly criticized his employer's deal on X, calling the safeguards "window dressing."
(Katrina Mulligan, LinkedIn, February 2026 / Leo Gao, X, February 2026 / TechCrunch, February 28, 2026)
Altman later said: "I think this is a very bad decision from the DoW, and I hope they reverse it. If we take heat for strongly criticizing it, so be it." He was referring to the Anthropic designation.
(Sam Altman, X, March 2026 / TechCrunch, March 2026)
The difference between the two deals, by Altman's own account, was that OpenAI was willing to move when asked. Anthropic was not.
During his CBS News interview, Amodei used one phrase consistently when referring to the Pentagon: the Department of War. He used it throughout the interview, not as a slip, but as the name he had internalized for the institution he had been negotiating with. The Trump administration began using that name in January. It is not the department's legal name. Congress has not passed legislation changing it. It is, so far, a branding preference enforced through use.
Amodei used it nine times.
(CBS News full transcript, March 2026)
The two restrictions Anthropic refused to remove are worth understanding in specific terms, because they describe capabilities that did not exist at meaningful scale before this era of AI.
The first is domestic mass surveillance. Amodei described the mechanism in his CBS interview: private firms collect data on Americans — location, personal associations, political affiliation — and sell it to the government. That transaction is currently legal. Before AI, analyzing that data at scale was not practically useful. It is now. The Fourth Amendment's judicial interpretations were written before this capability existed. As Amodei put it: "The technology is advancing so fast that it's out of step with the law."
The second is fully autonomous weapons — systems that target and fire without a human in the decision loop. Amodei was not categorically opposed to the concept; he acknowledged the US may eventually need them to match adversaries. His objection was technical and structural: AI systems today are not reliable enough for that use, and no oversight framework exists. He offered to work with the Pentagon to prototype such systems in a sandbox. He said they declined unless they could operate without restrictions from the start.
Senators on the Armed Services Committee wrote to Hegseth and Amodei urging them to extend negotiations and work with Congress. Their letter, obtained by ABC News, acknowledged that the question of what constitutes "lawful use" in the age of AI "requires additional work by all stakeholders" and that legislative or regulatory language may be needed. Congress, by Amodei's own assessment, "doesn't move fast."
(ABC News, March 2026 / CBS News full transcript, March 2026)
The pattern is not primarily a dispute between a tech company and the Pentagon. It is a demonstration of what governing instruments now look like in practice.
A war was announced on Truth Social at 2:30 AM. A company's federal contracts were terminated via a post on X. A supply chain designation — with claimed scope that exceeds what the authorizing statute allows — was delivered through social media before any formal legal action was taken. Compliance followed the posts, not the filings.
Anthropic's statement described the designation as setting "a dangerous precedent for any American company that negotiates with the government." The Senate letter described a Congress that recognizes its own inadequacy on the underlying technology questions.
Congress voted on war powers resolutions in both chambers this week. Both failed. The House vote was 212-219. The Senate vote was 47-53. They were the seventh and eighth consecutive war powers resolutions to fail since June 2025. All eight have failed.
(ABC News, March 4-5, 2026 / JURIST, March 5, 2026 / Washington Post, March 4, 2026)
The formal designation letter arrived five days after the tweets. The court filing came after that. The war powers votes came after the war started. The oversight mechanisms are present. They are operating after the fact, on a timeline that does not match the speed of the instruments they are meant to check.
The instruments are tweets. The effects are real. The law arrives later, when it arrives at all.
The historical doctrine connecting this sequence to prior US interventions is in Mask Off.