OpenAI's Pentagon Deal Draws Scrutiny as Safety Claims Unravel

Within hours of Sam Altman posting that OpenAI's new Pentagon contract included safety provisions comparable to those Anthropic demanded, Under Secretary of State Jeremy Lewin contradicted him publicly — on X — and the gap between Silicon Valley's safety rhetoric and Washington's operational reality snapped into view.

The OpenAI contract flows from the principle of "all lawful use" that the Department of Defense insisted upon, Lewin wrote Feb. 28. The contrast between OpenAI's approach and Anthropic's refusal to deploy its models for autonomous weapons or domestic surveillance could not be starker.

The Pentagon had designated Anthropic a supply chain risk to national security after CEO Dario Amodei refused to allow its Claude models to be used for mass surveillance or automated killing systems. President Trump ordered all federal agencies to immediately halt use of Anthropic technology the same day. That decision created a vacuum, and Altman moved quickly to fill it.

But the safety pretense had already collapsed weeks earlier. Anthropic quietly released its Responsible Scaling Policy version 3.0 on Feb. 25, removing its pledge to pause AI training when safety measures fell short. The company replaced binding commitments with nonbinding goals it would "openly grade our progress towards" — a retreat that did not go unnoticed.

"We felt that it wouldn't actually help anyone for us to stop training AI models," said chief science officer Jared Kaplan. "We didn't really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments if competitors are blazing ahead."

Max Tegmark, founder of the Future of Life Institute, placed the blame squarely on the industry. "It's their fault that we have the race condition in the first place," he said. "All of the AI labs succumb to the incentives. It's just maybe Anthropic is the most striking one because they were the ones who always talk such a big game about safety."

The warning had come from inside. Mrinank Sharma, lead of Anthropic's Safeguards Research Team, resigned Feb. 9 and published a letter that cut to the bone: "The world is in peril. Throughout my time here, I've repeatedly seen how hard it is to truly let our values govern our actions. I've seen this within myself, within the organization, where we constantly face pressures to set aside what matters most."

The OpenAI contract, announced Feb. 27, states the Department of Defense may use the AI system for all lawful purposes. One provision bars the system from intentional use for domestic surveillance of U.S. persons and nationals, citing applicable laws including the Fourth Amendment. Legal experts are skeptical that the language delivers what it promises.

"I don't believe that provision is in the contract. I say that reluctantly, but I don't," said Carson. "The word surveillance doesn't even include the kind of activities that people are most concerned about. They're trying to blind you with complicated legal terms that ordinary people think mean something different entirely. But the lawyers know what it means. And the lawyers know that this is no guardrail at all."

Lewin defended the contract's structure, arguing it appropriately vests legal questions in the democratic system — and contrasting that with Anthropic's approach, which he described as "unacceptably vesting those questions in a single unaccountable CEO who would usurp sovereign control of our most sensitive systems."

The Pentagon's treatment of Anthropic, however, signals where the real leverage lies. "By blacklisting Anthropic, the Defense Department demonstrated that if it comes to an impasse, they are not afraid to place extreme sanctions on a private U.S. company," said Jessica Tillipman, a government contracts law expert at George Washington University. The actual red line is apparent: AI firms work on the government's terms, or not at all.

For those watching AI's trajectory, the stakes extend well beyond any single contract. "This is more evidence that society is not prepared for the potential catastrophic risks posed by AI," said Chris Painter, director of policy at METR. "This is showing Anthropic believes it needs to shift into triage mode with its safety plans, because methods to assess and mitigate risk are not keeping up with the pace of capabilities."

Sharma's resignation letter, written from inside one of the world's most safety-focused AI labs, may be the most unsettling document in this story. If even Anthropic's own safeguards chief concluded that commercial pressure routinely overrides stated values, the question is no longer whether the industry can self-regulate — it is whether anyone is left who believes it can.