The load-bearing phrase was "all lawful uses"

Anthropic and the Pentagon spent months negotiating a contract that collapsed on three words. Court filings unsealed this week in the Northern District of California — 346 pages of emails between Dario Amodei and Pentagon undersecretary Emil Michael — show the deal died not on capability but on contract form. The Department of Defense wanted a licence for "all lawful uses." Anthropic wanted two hard exclusions: fully autonomous lethal weapons, and mass domestic surveillance of Americans. Those things do not fit together, and now Anthropic sits under a supply-chain risk designation with 180-day contract cancellations running while the D.C. Circuit hears its challenge.¹

What actually broke. The negotiation did not fail because Claude cannot do what DoD wanted Claude to do. It failed because the proposed contractual language would have moved the locus of permission. Anthropic's Responsible Scaling Policy — RSP, the internal governance framework that defines what the company will and will not let its models be used for — sets vendor-side red lines. "All lawful uses" sets buyer-side red lines, bounded only by statutory legality as interpreted by DoD and DOJ counsel. Autonomous weapons systems and bulk domestic surveillance can be constructed as lawful under existing U.S. authorities. So the DoD framing did not soften Anthropic's constraints; it dissolved them.

This is a standard government-contracting construct on the DoD side. Federal procurement typically conditions vendor obligations on statutory legality, not on vendor policy preferences. So one honest reading is that Anthropic's RSP is simply incompatible with the standard form the Pentagon writes, and this was a procurement fit problem dressed up as a values dispute. That reading is available. But it does not survive contact with what Michael's team put in writing.

The emails do not read like a procurement fit dispute. Reported excerpts show senior defense officials characterising Anthropic's conditions as ideologically "woke" and incompatible with machine-speed targeting concepts.² That is not the register of counsel negotiating around a policy incompatibility. It is the register of a buyer who wants the vendor to fold. Anthropic did not fold, and the supply-chain risk designation followed.

Then there is Emil Michael's xAI position. The unsealed documents show the Pentagon official who led negotiations held undisclosed equity in xAI — Elon Musk's frontier lab, an Anthropic competitor actively pursuing DoD and intelligence-community contracts — during the negotiation.¹ Take that seriously for a moment. The government official deciding whether to designate Anthropic a supply-chain risk had a financial interest in a company that stands to absorb the contract surface Anthropic vacates. This is not yet adjudicated as a material conflict, and the designation may have independent grounds. But the designation is now a contested market outcome, not a neutral regulatory finding, and the D.C. Circuit will hear it that way.

What Anthropic actually lost. The supply-chain risk designation triggers automatic 180-day wind-down on existing DoD contracts and blocks new prime contracting work.³ It is a bounded exclusion, not a permanent one. If Anthropic wins at the D.C. Circuit, the designation vacates and the market reopens. If it does not, xAI, Palantir, Cohere and other DoD-cleared frontier vendors absorb the pipeline for the duration.

The revenue hit is probably less catastrophic than the headlines suggest. The frontier-model federal market is currently dominated by Microsoft/OpenAI through FedRAMP-approved Azure OpenAI, and Anthropic reaches government buyers through AWS GovCloud and adjacent channels that do not depend on DoD prime status. The exclusion is a positioning wound, not a solvency event. But positioning is what this negotiation was about.

The RSP is now a public spec. Before the unsealing, Anthropic's red lines on autonomous weapons and domestic surveillance were policy statements on a company website. After the unsealing, they are legally-legible commitments made in writing to the U.S. government and reproduced in a federal court docket. Every future enterprise and government buyer now knows exactly where Anthropic will and will not go. Amodei has, in effect, published the operating spec of his safety posture — not by choice, but the outcome is the same.

This is the AI safety as market position frame doing its work. Anthropic has spent two years arguing that safety constraints are a feature for buyers who care about liability, accountability, and the ability to say in a deposition that they used the model with the published red lines. The DoD outcome is the first hard test of whether that positioning has a government buyer at scale. The answer, provisionally, is: not at the Pentagon, not under Emil Michael, not on "all lawful uses" terms. Whether it has one anywhere else — allied defence ministries, intelligence-community customers with narrower use profiles, civilian federal agencies — is the open question.

The gap the filings do not close. The unsealed documents describe the negotiation. They do not describe the model. If Claude can be fine-tuned or prompted into behaviour that approaches autonomous targeting or bulk surveillance analytics, then the contractual red line and the technical red line are not the same line, and Anthropic's positioning claim rests on governance rather than architecture. Safety as market position works only when the technical constraint is verifiable. Nothing in the 346 pages verifies it. That is not a criticism of Anthropic; it is a note on what this evidence can and cannot support.

The market read, then. Anthropic has traded near-term DoD prime revenue for a public commitment to constraints it was already making privately. xAI has, at minimum, a friendly procurement environment and, at maximum, a competitor sidelined by an official with equity in the outcome. The D.C. Circuit will decide whether that sequence is lawful. The commercial question, whether safety-as-position pays anywhere the Pentagon does not shop, is decided by which buyers show up next.

Footnotes

¶