OpenAI and Pentagon Deny Controversial Clause in Military AI Contract Amidst FOIA Battle
September 8, 2026
OpenAI’s Pentagon agreement is framed to enable independent verification that safety boundaries are respected and to prevent unrestricted Pentagon access to its models, presenting safeguards as part of democratic oversight of advanced AI.
The clause in question appeared in an updated contract version labeled P00003, tied to a potential 200 million dollar, two-year deal expanding OpenAI’s wartime mission model work for national security use cases.
The released text outlines eight prototype tasks, including identifying defense use cases for frontier AI, policy prototyping, risk forecasting, industry information sharing, and hands-on prototyping in intelligence analysis, cybersecurity, and autonomous systems, with a ChatGPT Gov application deployed on GenAI.mil at Impact Level 5.
OpenAI and the Pentagon deny that the final executed contract includes the ‘minimal refusal rates’ language, saying the document provided was a draft and not part of the final agreement.
Conflicting narratives have emerged: OpenAI disputes that the minimal-refusal-rate language was agreed or appears in the executed contract, while the Pentagon asserts it does not appear in any active contract; The Intercept reports ongoing FOIA litigation to obtain final contract documents.
Both sides maintain that the final contract did not contain language mandating minimal refusal rates or other overly restrictive terms.
A leaked Defense Department modification to an OpenAI prototype agreement defines OpenAI Mission Models as those designed for national security uses with minimal refusal rates, applicable to testing and refinement under Task 7 of the Frontier AI Model Prototype effort.
Anthropic’s legal battles and government actions set a backdrop for how restrictions and safety classifications intersect with private AI developers in defense contexts.
OpenAI says it won red lines on autonomous killings and spying on Americans, but contract language may still allow government-approved uses under the current framework.
Experts warn that ‘minimal refusal’ could signal looser safeguards and greater influence of private sector on warfare policy, potentially easing safety constraints.
OpenAI signed a February agreement allowing services to operate across the DoD’s classified networks, while the current contract text remains redacted in the latest deployment update.
OpenAI emphasizes three red lines—no mass domestic surveillance, no autonomous weapons without human control, and no high-stakes automated decisions—while noting the deployment is cloud-based with safety controls retained; a March 2026 update adds a domestic-surveillance restriction and says NSA-level intelligence would require a new agreement.
Summary based on 4 sources
Get a daily email with more Tech stories
Sources

The Intercept • Sep 8, 2026
The Pentagon Asked OpenAI for Artificial Intelligence Designed to Rarely Say No
Unite.AI • Sep 8, 2026
OpenAI Pentagon Contract Defines ‘Mission Models’ by Minimal Refusal Rates