Unsealed Emails Show the Pentagon Told Anthropic Its Ethics Were 'Just Not Workable'

On July 2, a federal court in Northern California unsealed a batch of emails that finally answer a question people have been asking since March: what actually broke down between Anthropic and the Pentagon before the Department of Defense labeled the company a "supply-chain risk to national security"? The answer, in Emil Michael's own words, is that Anthropic's ethical limits were "just not workable." Michael is the Pentagon's Under Secretary of Defense for Research and Engineering, and the recipient of his email was Anthropic CEO Dario Amodei, at the tail end of negotiations over deploying Claude on the DOD's GenAI.mil platform. The line reads less like a negotiating position and more like a verdict.

Two Rules, One Collision

Anthropic has exactly two hard limits baked into how it will let Claude be used by any customer, including the US government: no fully autonomous weapons systems that can engage a target without a human in the loop at the moment of firing, and no domestic mass surveillance. Those aren't new — Anthropic has said as much publicly for years. What the unsealed emails show is what happened when those limits collided with an actual $200 million contract, signed in July 2025, that both sides needed to work in practice. As negotiations over full deployment continued into September, Amodei told Michael the Pentagon's proposed contract language appeared to "completely remove our redlines." Michael didn't deny it. He confirmed it, and told Amodei there was "one more chance to align on core principles" before the two sides would have to part ways.

The Timing Is the Story

The emails matter less for what they say about ideology and more for what they establish about sequence. The day after Michael's "just not workable" email, Defense Secretary Pete Hegseth's department designated Anthropic a supply-chain risk — a label that bars defense contractors, suppliers, and partners from working with the company at all, with a six-month transition window for anyone still using Claude. Anthropic has maintained since March that the designation was retaliation dressed up as a security finding. The Pentagon has maintained it was a legitimate risk assessment. Emails showing the department's own point man calling the negotiation "very close" one day and torching it the next make Anthropic's version considerably easier to believe, regardless of what a judge ultimately rules.

The Courts Are Already Split

This isn't a hypothetical fight — it's already produced conflicting rulings. In April, the DC Circuit denied Anthropic's request to block the blacklisting while litigation continues, reasoning that the government's interest in "securing vital AI technology during an active military conflict" outweighed the financial harm to a single company. A separate federal judge in San Francisco reached the opposite conclusion in a related case, granting Anthropic a preliminary injunction against enforcement of a broader Claude ban across other government agencies. The net effect: Anthropic is currently locked out of new DOD contracts specifically, but can still sell to the rest of the federal government while the underlying question — can a national-security customer punish a vendor for refusing to remove its own ethical constraints — works its way through appeals.

Why This Outlasts One Contract

The Anthropic-Pentagon fight is being watched closely by every AI lab with defense ambitions, because it's the first real test of whether a company's safety commitments survive contact with its largest possible customer. OpenAI, Google DeepMind, and Palantir have all pursued defense contracts without publicly drawing lines as specific as Anthropic's — no fully autonomous weapons, no domestic surveillance — which made Anthropic the test case almost by default. If the DOD can successfully use a supply-chain risk designation as leverage to override a vendor's stated ethical limits, that's the playbook for every future negotiation with every future AI vendor. If Anthropic wins on appeal, it sets a precedent that a company can hold a specific line against the Pentagon and survive the retaliation. Either outcome will shape how the next AI-defense contract gets negotiated — this time with the emails from the last one already public.