OpenAI Pentagon Contract Defines ‘Mission Models’ by Minimal Refusal Rates

A Defense Department modification to its prototype agreement with OpenAI, released through a Freedom of Information Act lawsuit, defines the “OpenAI Mission Models” to be tested for the military as models “designed for national security use cases” that “have minimal refusal rates.” The language appears in the P00003 modification, the released agreement text, which was produced to The Intercept in litigation brought by Legal Advocates for Safe Science and Technology, the legal group representing the outlet’s FOIA requests.

OpenAI and the Pentagon dispute what the document represents. OpenAI spokesperson Nate Evans said the company never agreed to contract language requiring minimal refusal rates and that the phrase does not appear in the executed contract, describing the produced document as an earlier draft the Department proposed before OpenAI rejected the language. Defense Department spokesperson Jacob Bliss said the phrase does not appear in any active Department of War contract with OpenAI.

What the Released Modification Contains

The P00003 text modifies an Other Transaction Agreement between OpenAI Public Sector, LLC and the Defense Department’s Chief Digital and AI Office, the buyer for what the document calls the Chief Digital Artificial Intelligence Office Frontier AI Model Prototype effort. The agreement is a prototype project executed under the Pentagon’s prototype-transaction authority, with a project period running from June 13, 2025 to June 12, 2027.

Task 7 of the prototype work statement is “Testing, Evaluation, and Refinement of OpenAI Mission Models.” Its description says the task supports collaborative testing, evaluation, and refinement of the mission models on national security problem sets, and it defines those models as OpenAI models designed for national security use cases with minimal refusal rates. The activities and deliverables beneath the task are redacted under withholding exemptions, as is much of the surrounding technical material.

The modification’s separately released signature form is dated January 30, 2026, with a contractor signature date of February 6, 2026. The form states that the modification establishes and funds a new contract line item for the Frontier AI Model prototype, incorporates “Prototype Agreement HQ0883-25-9-0012 Version 1.0” as an attachment, and increases the total obligated amount by $1, from $1,999,998 to $1,999,999, with a line-item period of performance from January 30, 2026 to June 12, 2027.

Elsewhere, the agreement states a ceiling at award of $200,000,000, with payment milestones at monthly increments and a provision allowing follow-on production without competitive procedures if the prototype is successfully completed.

The Prototype Tasks Around It

The released text lays out eight prototype tasks. They include identifying and scoping defense use cases for frontier AI, policy prototyping and validation exercises, risk forecasting and threat ideation, information sharing between industry and the Department, administrative and programmatic support, and hands-on prototyping and solutions engineering in areas the document lists as intelligence analysis, cybersecurity, and autonomous systems.

One task covers deploying a ChatGPT Gov application on GenAI.mil, described as the Department’s secure generative AI portal, at Impact Level 5 security requirements. The prototype agreement states the objective of identifying defense operational workflows in which frontier AI may deliver military advantage, and success criteria benchmarked on validation by end users such as warfighters and commanders. The released version also provides for forward deployed engineers to aid government employees in warfighting support settings, including combatant commands, service components, and theatre components.

The document quotes Secretary Pete Hegseth describing artificial intelligence as a set of capabilities the Department must use, citing adversary investment in such capabilities, particularly China.

The Later Version and the Disclosure Dispute

A subsequent modification, P00004, effective February 27, 2026 and signed by both parties that day, incorporates a Version 2.0 of the prototype agreement. In that version, the equivalent task is renamed “Testing, Evaluation, and Refinement of ChatGPT Mission Models,” and its entire description is redacted, so the released text does not show whether a refusal-rate definition remains. The later version also changes the GenAI.mil deployment task from ChatGPT Gov to “ChatGPT Mil” and directs the Department and OpenAI to jointly engage Microsoft within 30 days to negotiate in good faith a reduction or waiver of Azure cloud compute costs.

The disclosure has produced conflicting accounts of the P00003 document’s status. The FOIA request underlying the lawsuit sought only final, executed contract documents and asked the Pentagon to exclude drafts; none of the produced documents is marked as a draft. According to The Intercept’s account of the litigation, a Justice Department attorney representing the Pentagon first confirmed the document was the signed and executed version, then withdrew that confirmation and said the Department needed more time to investigate, and later said the document “was not the final version” and that a correct document would be shared without a definitive timeline. The Intercept’s FOIA counsel, Legal Advocates for Safe Science and Technology, sued the Pentagon in April 2026 in the U.S. District Court for the Southern District of New York after the Department missed the statutory deadline on the group’s March 5, 2026 requests for the AI company contracts.

OpenAI’s Stated Contract Terms

OpenAI has published its own account of the classified-deployment agreement reached February 27, 2026. In its statement on the Department of War agreement, the company lists three red lines: no use of its technology for mass domestic surveillance, no use to direct autonomous weapons systems, and no use for high-stakes automated decisions such as social credit systems. The company says the deployment is cloud-only, that it retains full discretion over its safety stack, and that it is not providing “guardrails off” or non-safety-trained models.

The contract language OpenAI published states that the Department may use the AI system for all lawful purposes and that the system will not be used to independently direct autonomous weapons where law, regulation, or Department policy requires human control. An update dated March 2, 2026 adds language stating the system shall not be intentionally used for domestic surveillance of U.S. persons and nationals, including through commercially acquired personal or identifiable information, and affirms that Department of War intelligence agencies such as the NSA would require a new agreement for any services.