Why the US Defense Department Blacklisted Anthropic: Inside the AI Power Struggle

Background and Contract Dispute

In July 2025 Anthropic signed a $200 million contract with the U.S. Department of Defense, deploying its Claude model on the Pentagon’s classified network. The agreement included a strict acceptable‑use policy prohibiting large‑scale surveillance of U.S. citizens and the use of the model in fully autonomous weapon systems.

After the model was used via Palantir’s API to process intelligence for a Venezuelan operation that attempted to capture former President Maduro, Anthropic asked Palantir for details on Claude’s specific role. The Pentagon deemed this inquiry an overreach, asserting that once a tool is sold to the military the supplier has no right to audit its downstream use.

Pentagon Ultimatum and “All Legitimate Use” Clause

In early 2026 the Pentagon introduced a new standard contract term—dubbed “All Legitimate Use”—which obliges AI suppliers to comply with any use the military deems lawful, without the supplier setting additional limits. OpenAI, Google, and xAI accepted the clause; Anthropic refused, arguing that the term is too vague and could permit prohibited applications.

The Pentagon warned that refusal would trigger contract termination, a supply‑chain risk label, and possible invocation of the Defense Production Act.

Industry Reaction and Internal Splits

The dispute polarized Silicon Valley. Elon Musk’s xAI publicly aligned with the Pentagon, adopting the “All Legitimate Use” terms, while many Google engineers petitioned for similar restrictions on Gemini. Over 100 Google staff signed a letter to Jeff Dean demanding red‑line constraints. Microsoft and Amazon employees raised comparable concerns.

Anthropic’s co‑founder Dario Amodei publicly refused to compromise, stating the contract’s restrictions were essential for safety. OpenAI’s Sam Altman later announced a separate defense contract, deploying OpenAI models under the same “legitimate use” framework, which ignited further debate about the adequacy of the safeguards.

Legal Ambiguities and Future Outlook

The “All Legitimate Use” language is legally ambiguous; it allows the Department of Defense to employ the AI system for any purpose that complies with applicable law, potentially eroding the original red‑line prohibitions. Anthropic plans to challenge the supply‑chain risk label in court, citing 10 U.S.C. § 3252, which limits the government’s authority to dictate a contractor’s use of technology beyond the scope of the specific contract.

With a six‑month transition period ending in August 2026, Anthropic must continue providing services while the Pentagon seeks alternative models, such as the less‑advanced Grok system. Meanwhile, OpenAI’s agreement includes defensive technical measures: cloud‑only deployment, exclusion of edge devices, and monitoring by cleared engineers.

Conclusion

The clash illustrates a fundamental tension between frontier AI developers seeking ethical safeguards and a national security apparatus demanding unrestricted operational control. As AI models become integral to military decision‑making, the definition of “legitimate use” will likely become a focal point of future policy and legal battles.