OpenAI Discloses Further Terms of Pentagon Partnership

OpenAI CEO Sam Altman has acknowledged that the company's agreement with the Department of Defense was finalized too quickly and that the public perception is unfavorable.

Following the breakdown of talks between Anthropic and the Pentagon on Friday, President Donald Trump ordered federal agencies to stop using Anthropic's technology after a six-month transition. Defense Secretary Pete Hegseth labeled the AI firm a supply-chain risk.

OpenAI then promptly announced its own agreement to deploy models in classified settings. With Anthropic stating it would prohibit its technology from being used in fully autonomous weapons or mass domestic surveillance—and Altman affirming OpenAI holds the same restrictions—questions arose: Was OpenAI being transparent about its safeguards? Why did it secure a deal when Anthropic could not?

As OpenAI executives defended the arrangement on social media, the company also released a blog post detailing its strategy.

The post highlighted three specific applications where OpenAI's models are not permitted: mass domestic surveillance, autonomous weapon systems, and "high-stakes automated decisions (such as 'social credit' systems)."

OpenAI stated that, unlike other AI companies that have "weakened or removed safety measures and depend mainly on usage policies as safeguards in national security deployments," its agreement upholds its red lines "through a broader, multi-layered approach."

"We maintain full control over our safety infrastructure, deploy via cloud, have authorized OpenAI personnel involved, and enforce strong contractual protections," the blog explained. "These are in addition to the robust protections already provided by U.S. law."

The company further noted, "We are unsure why Anthropic was unable to reach a similar agreement, and we hope they and other labs will consider this path."

After the post went live, Techdirt's Mike Masnick contended that the deal "does permit domestic surveillance," since it states that private data collection will follow Executive Order 12333 and other laws. Masnick characterized that order as "the legal basis the NSA uses to conduct domestic surveillance by intercepting communications outside the U.S., even when they involve U.S. citizens."

In a LinkedIn post, Katrina Mulligan, OpenAI's head of national security partnerships, argued that much of the debate about the contract language assumes "the only barrier against using AI for mass domestic surveillance and autonomous weapons is a single clause in one contract with the Department of War."

"That is not how the system operates," Mulligan stated, adding, "Deployment architecture is more critical than contract wording. By restricting our deployment to cloud API, we can prevent our models from being directly integrated into weapons, sensors, or other operational hardware."

Altman also addressed questions about the agreement on X, admitting it was rushed and provoked significant backlash against OpenAI—so much so that Anthropic's Claude surpassed OpenAI's ChatGPT on the Apple App Store that Saturday. So why proceed?

"We genuinely sought to de-escalate tensions and believed the proposed deal was beneficial," Altman said. "If we are correct and this leads to reduced friction between the Department of War and the industry, we will be seen as visionary—a company that endured considerable difficulty to support the sector. If not, we will continue to be criticized for acting hastily and without due care."