OpenAI CEO Sam Altman continues to be within the sizzling seat this week after his firm signed a cope with the US army. OpenAI workers have criticized the transfer, which got here after Anthropic’s roughly $200 million contract with the Pentagon imploded, and requested Altman to launch extra details about the settlement. Altman admitted it appeared “sloppy” in a social media post.
Whereas this incident has grow to be a serious information story, it could simply be the newest and most public instance of OpenAI creating obscure insurance policies round how the US army can entry its AI.
In 2023, OpenAI’s utilization coverage explicitly banned the army from accessing its AI fashions. However some OpenAI workers found the Pentagon had already began experimenting with Azure OpenAI, a model of OpenAI’s fashions provided by Microsoft, two sources accustomed to the matter mentioned. On the time, Microsoft had been contracting with the Division of Protection for many years. It was additionally OpenAI’s largest investor, and had broad license to commercialize the startup’s know-how.
That very same 12 months, OpenAI workers noticed Pentagon officers strolling via the corporate’s San Francisco workplaces, the sources mentioned. They spoke on the situation of anonymity as they aren’t licensed to touch upon personal firm issues.
Some OpenAI workers had been cautious about associating with the Pentagon, whereas others had been merely confused about what OpenAI’s utilization insurance policies meant. Did the coverage apply to Microsoft? Whereas sources inform WIRED it was not clear to most workers on the time, spokespeople from OpenAI and Microsoft say Azure OpenAI merchandise are usually not, and weren’t, topic to OpenAI’s insurance policies.
“Microsoft has a product known as the Azure OpenAI Service that turned accessible to the US Authorities in 2023 and is topic to Microsoft phrases of service,” mentioned spokesperson Frank Shaw in a press release to WIRED. Microsoft declined to remark particularly on when it made Azure OpenAI accessible to the Pentagon, however notes the service was not accepted for “top secret” authorities workloads till 2025.
“AI is already taking part in a big function in nationwide safety and we imagine it’s vital to have a seat on the desk to assist guarantee it’s deployed safely and responsibly,” OpenAI spokesperson Liz Bourgeois mentioned in a press release. “We have been clear with our workers as we’ve approached this work, offering common updates and devoted channels the place groups can ask questions and have interaction instantly with our nationwide safety group.”
The Division of Protection didn’t reply to WIRED’s request for remark.
By January 2024, OpenAI up to date its insurance policies to take away the blanket ban on army use. A number of OpenAI workers discovered concerning the coverage replace via an article in The Intercept, sources say. Firm leaders later addressed the change at an all-hands assembly, explaining how the corporate would tread rigorously on this space shifting ahead.
In December 2024, OpenAI introduced a partnership with Anduril to develop and deploy AI methods for “nationwide safety missions.” Forward of the announcement, OpenAI instructed workers that the partnership was slim in scope and would solely cope with unclassified workloads, the identical sources mentioned. This stood in distinction to a deal Anthropic had signed with Palantir, which might see Anthropic’s AI used for categorised army work.
Palantir approached OpenAI within the fall of 2024 to debate taking part of their “FedStart” program, an OpenAI spokesperson confirmed to WIRED. The corporate in the end turned it down, and instructed workers it could’ve been too high-risk, two sources accustomed to the matter inform WIRED. Nevertheless, OpenAI now works with Palantir in different methods.
Across the time the Anduril deal was introduced, a couple of dozen OpenAI workers joined a public Slack channel to debate their issues concerning the firm’s army partnerships, sources say and a spokesperson confirmed. Some believed the corporate’s fashions had been too unreliable to deal with a person’s bank card info, not to mention help People on the battlefield.
