The Pentagon once asked OpenAI for a military edition of its AI designed to refuse commands as rarely as possible, and the two sides now disagree over whether that request ever became binding.
The clause appears in version P00003 of an agreement that expanded a prototype deal from last summer, worth up to $200M over two years, between OpenAI Public Sector and the Defense Department’s Chief Digital and AI Office. The outlet pried the papers loose through a FOIA lawsuit filed jointly with the Legal Advocates for Safe Science and Technology.
OpenAI spokesman Nate Evans said his company never agreed to such language and cast the released text as a draft the Department put forward before OpenAI pushed back. Jacob Bliss, speaking for the Pentagon, said no current OpenAI contract includes the language.
Other details complicate that account. The FOIA request was limited to final, executed documents, and none of what came out is marked as a draft. A Justice Department lawyer first told The Intercept the P00003 file was the signed version, then walked that back hours later, after the outlet contacted OpenAI, saying the Department needed more time to look into the matter.
OpenAI was one of four frontier labs, alongside Google, xAI and Anthropic, that agreed in 2025 to build militarized prototypes for logistics, intelligence and warfighting support. The refusal-rate clash echoes a separate fight this year over Pentagon plans to put Anthropic’s models on classified networks.
