Anthropic Loses to the Pentagon: Aligning Claude Counts as "Manipulating" It

Key takeaways

1. 1

   The D.C. Circuit Court of Appeals (Judges Katsas, Rao, and Henderson) rejects Anthropic's challenges 2-1; Judge Karen LeCraft Henderson writes a dissenting opinion.

2. 2

   Anthropic had agreed to significantly loosen its restrictions (weapons design, foreign intelligence, offensive cyber), but refused to let Claude be used for autonomous lethal warfare and mass surveillance of Americans.

3. 3

   On March 3, 2026, Pete Hegseth excluded Claude under the 2018 Federal Acquisition Supply Chain Security Act, requiring removal of systems within 180 days at most and barring subcontractors.

4. 4

   For the majority, "manipulating" a product's design or operation requires no malicious intent: the risk definition "turns on what Anthropic does, not why Anthropic does it."

5. 5

   Anthropic's technical argument (no "kill switch," frozen models once delivered, possible pre-testing) is dismissed: the Pentagon cannot make do with systems "trapped in amber" while models keep advancing.

6. 6

   The constitutional claims fail: post-hoc process is sufficient given the urgency, and the exclusion stems from a contractual disagreement, not punishment of Anthropic's speech about AI regulation.

7. 7

   The court distinguishes itself from an August 27, 2026 California ruling that struck down a designation based on a different statute (10 U.S.C. § 3252), whose narrower definition targets an "adversary."