OpenAI’s Sam Altman announces Pentagon deal with ‘technical safeguards’

OpenAI CEO Sam Altman announced precocious connected Friday that his institution has reached an statement allowing the Department of Defense to usage its AI models successful the department’s classified network.

This follows a high-profile standoff betwixt the section — besides known nether the Trump medication arsenic the Department of War — and OpenAI’s rival Anthropic. The Pentagon pushed AI companies, including Anthropic, to let their models beryllium utilized “all lawful purposes,” portion Anthropic sought to gully a reddish enactment astir wide home surveillance and afloat autonomous weapons.

In a lengthy connection released Thursday, Anthropic CEO Dario Amodei said the institution “never raised objections to peculiar subject operations nor attempted to bounds usage of our exertion successful an ad hoc manner,” but helium argued that “in a constrictive acceptable of cases, we judge AI tin undermine, alternatively than defend, antiauthoritarian values.”

More than 60 OpenAI employees and 300 Google employees signed an unfastened missive this week asking their employers to enactment Anthropic’s position.

After Anthropic and the Pentagon failed to scope an agreement, President Donald Trump criticized the “Leftwing seed jobs astatine Anthropic” successful a societal media post that besides directed national agencies to halt utilizing the company’s products aft a six-month signifier retired period.

In a abstracted post, Secretary of Defense Pete Hegseth claimed Anthropic was trying to “seize veto powerfulness implicit the operational decisions of the United States military.” Hegseth besides said helium is designating Anthropic arsenic a supply-chain risk: “Effective immediately, nary contractor, supplier, oregon spouse that does concern with the United States subject whitethorn behaviour immoderate commercialized enactment with Anthropic.”

On Friday, Anthropic said it had “not yet received nonstop connection from the Department of War oregon the White House connected the presumption of our negotiations,” but insisted it would “challenge immoderate proviso concatenation hazard designation successful court.”

Surprisingly, Altman claimed successful a station connected X that OpenAI’s caller defence declaration includes protections addressing the aforesaid issues that became a flashpoint for Anthropic.

“Two of our astir important information principles are prohibitions connected home wide surveillance and quality work for the usage of force, including for autonomous limb systems,” Altman said. “The DoW agrees with these principles, reflects them successful instrumentality and policy, and we enactment them into our agreement.”

Altman said OpenAI “will physique method safeguards to guarantee our models behave arsenic they should, which the DoW besides wanted,” and it volition deploy engineers with the Pentagon “to assistance with our models and to guarantee their safety.”

“We are asking the DoW to connection these aforesaid presumption to each AI companies, which successful our sentiment we deliberation everyone should beryllium consenting to accept,” Altman added. “We person expressed our beardown tendency to spot things de-escalate distant from ineligible and governmental actions and towards tenable agreements.”

Fortune’s Sharon Goldman reports that Altman told OpenAI employees astatine an all-hands gathering that the authorities volition let the institution to physique its ain “safety stack” to forestall misuse, and that “if the exemplary refuses to bash a task, past the authorities would not unit OpenAI to marque it bash that task.”

Altman’s station came soon earlier quality broke that the U.S. and Israeli governments have begun bombing Iran, with Trump calling for the overthrow of the Iranian government.