346 Pages of Emails Show Exactly How Anthropic and the Pentagon Broke Down Over Autonomous Weapons

The court documents are 346 pages. They were unsealed in July 2026 as part of ongoing litigation between Anthropic and the US Department of Defense, and they contain something the months of public statements on both sides have carefully avoided: the actual words exchanged between Anthropic CEO Dario Amodei and Pentagon undersecretary Emil Michael as their relationship collapsed in real time.

The emails document, in granular detail, one of the most consequential negotiations in the history of AI governance — and a breakdown that resulted in President Trump ordering all federal agencies to cease use of Anthropic’s technology, and Defense Secretary Pete Hegseth designating Anthropic a national security supply-chain risk.

The Setup: A $200 Million Contract

In July 2025, the Department of Defense awarded contracts to Anthropic, Google, OpenAI, and xAI worth up to $200 million each, intended to accelerate the Pentagon’s adoption of frontier AI capabilities for national security purposes. The contracts were competitive and prestigious, and all four labs accepted.

For Anthropic, the contract came with an assumption: that the company’s existing usage policies — which prohibit Claude from being deployed for certain categories of harm — would be incorporated into the government agreement. Amodei considered this non-negotiable. The Pentagon, as the emails would reveal, considered it non-workable.

The Core Demand: “Any Lawful Use”

Renegotiations over the contract language began in late 2025 and broke down definitively in January and February 2026. The pivot point was a single contractual clause.

The Pentagon demanded that Anthropic accept language authorizing Claude for “any lawful use” — a phrase that, in Anthropic’s reading, would eliminate two existing prohibitions that the company had embedded in its commercial terms of service: the bar on using Claude to power fully autonomous weapons systems that make lethal targeting decisions without meaningful human authorization, and the bar on using Claude to conduct mass domestic surveillance of American citizens.

In a January 2026 email, Michael wrote to Amodei that he was “hoping that we are closer to engaging with your revised POV” — suggesting he had expected Anthropic to have reconsidered its position during weeks of silence. Amodei’s response was direct: the requested language “completely remove[d] our redlines.” He would not sign.

Michael’s reply escalated the stakes explicitly: this was “one more chance to align on core principles” before the Pentagon would “decide to part ways.”

The Blacklisting

When Amodei did not move, the consequences came quickly. Defense Secretary Pete Hegseth designated Anthropic a national security supply-chain risk. President Trump directed all federal agencies to immediately cease use of Anthropic’s technology. Defense contractors, suppliers, and partners were barred from working with the company.

For Anthropic, the designation had immediate practical consequences: the company’s ability to do business with the substantial US government AI market — estimated at billions of dollars annually — was effectively suspended.

The Contradiction That Defined the Case

What the newly unsealed documents make most starkly visible is a sequence that Anthropic’s legal team has argued is conclusive evidence of bad faith: the timeline gap between the blacklisting and the subsequent communication.

On March 4, 2026 — the day after the Pentagon formally finalized its supply-chain risk designation — Michael emailed Amodei to say the two sides were “very close” on the very issues the government had just cited as evidence that Anthropic posed a national security threat.

That email became central to Anthropic’s First Amendment lawsuit against the government. In March 2026, Judge Rita F. Lin granted a preliminary injunction against the Trump administration, finding that the Department of Defense’s records showed it had designated Anthropic a supply-chain risk primarily because of Anthropic’s “hostile manner through the press” — an assessment that, the judge concluded, constituted illegal retaliation for constitutionally protected speech.

The Conflict of Interest

The newly unsealed emails also draw attention to a detail that had previously received limited coverage: Emil Michael, the Pentagon undersecretary who conducted the negotiations on the government’s behalf, holds undisclosed financial interests in xAI — Elon Musk’s AI company and one of Anthropic’s direct competitors.

xAI was one of the four companies awarded the original July 2025 DoD contracts alongside Anthropic. With Anthropic effectively frozen out of government contracts, xAI — and its competitors at Google and OpenAI — would be the primary beneficiaries of any federal AI spending that would otherwise have flowed to Anthropic.

The potential conflict raises questions that neither Michael’s office nor the Pentagon has addressed publicly: whether the terms of the negotiation were structured in a way that any reasonable reading of Anthropic’s usage policies would make compliance impossible, and whether the standard applied to Anthropic was applied equally to the other contract holders.

What the Emails Reveal About AI Governance

Beyond the personalities and the litigation, the 346 pages of email reveal something more fundamental: the absence of a shared framework for what an AI company owes its government customers, and what limits it is permitted to maintain when those customers represent the most powerful military in the world.

Amodei’s position — that an AI lab can and should maintain ethical use restrictions that apply even to government contracts — was treated by Michael as operationally unacceptable. Michael’s position — that government users should not have their operational capabilities constrained by a private company’s ethics policy — was treated by Amodei as a non-starter.

Neither side was wrong, exactly, within its own frame. But the frames are irreconcilable, and the US government has no mechanism for resolving the conflict other than coercion.

The Anthropic case is being watched carefully by every other AI company with government contracts. The implicit lesson is stark: if a US AI lab maintains ethical use restrictions that the government finds inconvenient, the government will designate it a supply-chain risk and freeze it out of the federal market.

Whether that lesson deters future AI companies from holding similar lines — or whether Anthropic’s court victory on First Amendment grounds creates a precedent that protects them — will be one of the defining governance questions of the AI era.