CEO Sam Altman to OpenAI staff: You don't get a say in how the military uses our AI

OpenAI CEO Sam Altman had a blunt connection for his employees this week: erstwhile it comes to US subject operations, the institution simply does not get a vote. "So possibly you deliberation the Iran onslaught was bully and the Venezuela penetration was bad," Altman told unit astatine an all-hands gathering connected Tuesday, according to a partial transcript reviewed by CNBC.

"You don't get to measurement successful connected that."The gathering came 4 days aft Altman announced, precocious connected a Friday evening, that OpenAI had struck a woody with the Pentagon to deploy its AI models connected classified networks—a woody that landed conscionable hours aft rival Anthropic was formally blacklisted by the Department of Defense and hours earlier the US and Israel launched strikes connected Iran.

OpenAI's Pentagon woody drew contiguous backlash—inside and extracurricular the company

The timing could not person been much loaded. Anthropic had conscionable been designated a "supply concatenation hazard to nationalist security" by Defence Secretary Pete Hegseth—an unprecedented statement for an American company—after it refused to driblet guardrails against AI being utilized for wide home surveillance of Americans oregon afloat autonomous weapons.

OpenAI stepped successful astir immediately, announcing its ain classified deployment woody earlier the particulate had adjacent settled.The optics weren't great. Altman himself admitted arsenic much. "We shouldn't person rushed to get this retired connected Friday," helium said successful a station connected X implicit the weekend. "The issues are ace complex, and request wide communication." In the all-hands, helium acknowledged it looked "opportunistic and sloppy," according to WSJ reporting connected the meeting.

The backlash was real. Some OpenAI employees publically criticised the move. Dozens had, conscionable days earlier, signed an unfastened missive lasting successful solidarity with Anthropic's reddish lines. The AI information assemblage was alarmed. And critics pointed retired that OpenAI's declaration language, portion it included prohibitions connected home surveillance and autonomous weapons successful principle, yet deferred to the ineligible framework—a model that, post-Snowden, galore reason has already been stretched to accommodate wide surveillance programs similar PRISM.

Altman says Pentagon respects OpenAI's information stack—but operational calls beryllium to Hegseth

Still, Altman drew wide lines astatine Tuesday's meeting. He told employees that the Pentagon respects OpenAI's method expertise, wants input connected wherever its models are a bully fit, and has agreed to fto the institution physique and support the information stack it deems appropriate, according to a idiosyncratic acquainted with the substance who spoke to CNBC connected information of anonymity. Cleared OpenAI engineers volition beryllium embedded with authorities teams, and information researchers volition stay successful the loop.But Altman was arsenic wide that day-to-day subject decisions are not OpenAI's to make. Secretary Pete Hegseth runs those calls—not Sam Altman.He besides addressed a competitory world that fewer astatine the institution wanted to hear. "I judge we volition hopefully person the champion models that volition promote the authorities to beryllium consenting to enactment with us, adjacent if our information stack annoys them," Altman said. "But determination volition beryllium astatine slightest 1 different actor, which I presume volition beryllium xAI, which efficaciously volition accidental 'We'll bash immoderate you want.'"

Altman present eyeing NATO classified networks arsenic OpenAI doubles down connected defence

And OpenAI is already looking beyond the Pentagon. The WSJ reported that Altman told unit the institution is present exploring a declaration to deploy connected each NATO classified networks—a determination that would marque OpenAI a foundational AI supplier for the Western subject alliance. Apple received NATO clearance for its user devices conscionable past month, but a afloat classified deployment of frontier AI models would beryllium a antithetic proposition entirely.Meanwhile, Anthropic's Claude was reportedly utilized successful the Iran strikes implicit the play and successful the January cognition that resulted successful the seizure of ousted Venezuelan person Nicolás Maduro—suggesting the classified handoff from Anthropic to OpenAI and xAI is inactive precise overmuch a enactment successful progress.Altman has said helium reiterated to the Pentagon that Anthropic should not beryllium labelled a proviso concatenation risk, and that the aforesaid woody presumption should beryllium made disposable to each AI companies. Whether that connection leads anywhere—or whether the standoff ends successful court—remains to beryllium seen.