Anthropic Rejects Pentagon Demand to Loosen AI Safeguards and Risks Contract Loss
Key Highlights
Anthropic said it won’t loosen safeguards that block Claude from being used for autonomous targeting or large-scale domestic surveillance.
The Pentagon set a Friday deadline and warned Anthropic could lose defense work and face a “supply-chain risk” designation if it refuses.
The dispute raises the possibility that the Defense Production Act could be used as leverage as rival AI firms move to sign broader “lawful use” terms for classified deployments.
Anthropic CEO Dario Amodei announced the company will not accept Defense Department terms that could allow its Claude model to be used for autonomous targeting or broad domestic surveillance, escalating a rare public dispute between a frontier AI lab and the Pentagon.
In a post published Thursday, Amodei wrote the government’s position was contradictory, warning that Anthropic could be treated as a security risk while also pressing for wider military use of its technology. He said the company “cannot in good conscience” agree to language that would allow safeguards to be overridden.
The Pentagon responded with a 5:01 p.m. ET Friday deadline, saying it would end the relationship and designate Anthropic a supply-chain risk if the company does not grant access under an “all lawful purposes” standard. The Defense Department has also raised the possibility of invoking the Defense Production Act, which can compel private-sector support for national defense priorities.
The dispute puts an estimated $200 million contract at risk, but the supply-chain designation could have broader effects by forcing contractors and vendors to avoid Anthropic tools in defense workflows.
The standoff comes as competing AI firms push deeper into government deployments. xAI has moved to bring Grok into classified systems under terms closer to what the Pentagon is seeking, increasing pressure on Anthropic as agencies standardize procurement language.
Anthropic said it remains open to further discussion, but its conditions, including no autonomous targeting and no mass surveillance, are non-negotiable.
