Warren demands Pentagon AI contracts as Senate NDAA writes rules for military autonomy

Sen. Elizabeth Warren gave the Pentagon and seven AI vendors until July 20 to open their classified-network AI agreements, weeks after Senate defense authorizers wrote human-judgment requirements and a nuclear-launch ban for military AI into their draft policy bill.

Sen. Elizabeth Warren asked Google, SpaceX, Nvidia, Reflection AI, Microsoft, Amazon Web Services and Oracle on Monday night for the full text of their military AI contracts, NBC News reported. A companion letter went to Defense Secretary Pete Hegseth.

The letters, released Tuesday by her office, follow the Pentagon's May 1 announcement of agreements with eight AI companies to deploy their models on classified military networks. The department has published almost no information about the contractual guardrails governing those deployments, Warren wrote, and any standard permitting "all lawful use" leaves the door open to civil-liberties violations and civilian harm. She set a July 20 deadline for answers on which models run in classified systems, in what situations, and whether the agreements allow domestic mass surveillance or lethal autonomous weapons. Warren also questioned the award to Reflection AI, which has not publicly released a model, and asked for any communications between department officials and Donald Trump Jr. about it.

On June 10 the Senate Armed Services Committee finished marking up its fiscal 2027 National Defense Authorization Act with a regulatory framework for military AI and autonomous weapons, Arms Control Today detailed in its July/August issue. The draft requires "appropriate levels of human judgment" over the use of force and obliges designers to give operators the means to intervene or shut a system down. An incident repository would track failures and near misses. The text also bars AI from any "decision to initiate the launch or detonation of a nuclear weapon."

The same bill, S. 4784, would clamp down on the supply-chain-risk authority Hegseth used this year against Anthropic, Bloomberg Law noted, calling the provision lawmakers' first step in reaction to that fight. Anthropic had refused contract terms permitting "any lawful use" of its Claude model and sued; U.S. District Judge Rita Lin blocked the designation on March 26 as likely First Amendment retaliation, per court filings quoted by TechTimes.

The committee that drafted the rules also urged the Pentagon to "maximize uses" of the technology, per Arms Control Today, and the House bill, approved in committee June 5, defers more to the department. Warren's answers come due July 20; conference between the chambers decides how much of the framework becomes law.