OpenAI backtracks on Pentagon deal it signed hours after Anthropic's blacklisting
What's the deal? OpenAI signed an agreement on February 28 to let the PentagonDealroom has a profile for this one. Try Dealroom → deploy its AI models in classified systems — hours after the TrumpDealroom has a profile for this one. Try Dealroom → administration blacklisted rival AnthropicDealroom has a profile for this one. Try Dealroom → for refusing a similar deal.
By March 3, OpenAI chief executive Sam AltmanDealroom has a profile for this one. Try Dealroom → had publicly called the rollout "opportunistic and sloppy" and said the company was renegotiating to add explicit prohibitions on domestic surveillance of US citizens and bar intelligence agencies such as the National Security AgencyDealroom has a profile for this one. Try Dealroom → from using its services without a separate contract.
The sequence of events began when Anthropic demanded two hard limits: no mass surveillance of Americans and no lethal autonomous weapons — defined as AI systems that can select and strike targets without human oversight. The Pentagon refused. The Trump administration labelled Anthropic a supply-chain risk, a designation normally reserved for foreign companies with cybersecurity concerns, and ordered federal agencies to stop using Claude. OpenAI announced its deal the same day.
Why now? Critics immediately questioned how OpenAI had secured its stated red lines when Anthropic had not. The answer, according to sources familiar with the negotiations, is that it largely had not.
OpenAI's original agreement permitted "any lawful use" — language that, given the US government's history of interpreting surveillance laws broadly, would have allowed significant data collection on Americans.
Legal experts noted that the laws OpenAI cited as safeguards — including the Foreign Intelligence Surveillance Act of 1978 — had been used to justify sweeping surveillance programmes exposed by Edward Snowden in 2013.
Altman acknowledged the problem in an internal memo shared publicly. "The issues are super complex, and demand clear communication," he wrote. "We shouldn't have rushed to get this out on February 28."
What could go wrong? Even with the revised language, legal analysts say the new terms do not address autonomous weapons and leave significant grey areas. The updated contract bars "intentional" domestic surveillance — a qualifier that critics say creates room for interpretation.
Independent lawyers have not been given access to the full contract, and OpenAI employees have called for that review. One OpenAI research scientist posted publicly that he did not think "this deal was worth it," drawing nearly 500,000 views.
Anthropic, meanwhile, said it would challenge its supply-chain risk designation in court. The political and legal fallout remains unresolved.
The signal: The episode exposed a fault line running through the entire AI industry: how far safety commitments hold when government contracts and competitive pressure collide.
Anthropic's refusal generated a wave of public support — Claude reached number one on Apple's App Store, up from sixth — while OpenAI faced protests outside its San Francisco offices. Whether Anthropic's hard-line approach proves commercially sustainable or commercially ruinous will likely shape how every AI company negotiates with governments for years to come.
Altman himself framed it as "a good learning experience" — a phrase that may prove either prescient or grimly ironic, depending on what the renegotiated contract actually permits.
Sources:
OpenAI
Bloomberg
Reuters
Forbes
The New York Times
Fortune
The Guardian
BBC
NBC
CNBC
The Verge
CNN
Business Insider
J.V.
