Pentagon Sought AI With 'Minimal Refusal Rates' for National Security

A FOIA-obtained Pentagon draft described OpenAI models for national security use with 'minimal refusal rates,' The Intercept reported. OpenAI and the Pentagon said the language never appeared in an active contract, though a Justice Department lawyer initially confirmed it. Anthropic's exclusion from classified-network deals sharpened the debate over removing AI safeguards.

• Draft contract with a reported ceiling of up to $200M over two years cited 'minimal refusal rates'
• OpenAI said it rejected the language; Pentagon denies it appears in active contracts
• Anthropic was excluded from classified-network deals after seeking surveillance limits
• A DOJ lawyer first confirmed the phrase, then the department walked back that position