The week the three frontier providers showed how they'd behave in a room with the Pentagon

On Monday and Tuesday, TechCrunch reported that employees at OpenAI and Google were publicly backing Anthropic's legal fight with the Pentagon. By the end of the week, Fortune confirmed the count: 950 Google employees had signed an open letter asking the company to follow Anthropic's lead and refuse a DoD contract without similar guardrails. Anthropic's two redlines were narrow and specific: no use of Claude for domestic mass surveillance, no use for autonomous weapons.

The Pentagon wouldn't accept those limits. Relations between the administration and Anthropic collapsed, and a judge granted Anthropic an injunction against the contract designation while the case proceeds. OpenAI moved in the other direction in the same week, signing a DoD deal on terms the Pentagon found acceptable. Google followed, despite the open letter. Three providers, three postures, all crystallised inside seventy-two hours.

This is the part of the AI-safety conversation that does not show up in alignment papers. The technical question of whether a frontier model can be controlled is, in the end, less consequential than the contractual question of who gets to use it for what. Two of the three providers answered that question this week by accepting terms their own employees objected to. The third answered it by losing the contract.

Underneath the policy fight, OpenAI shipped two substantive product updates: improvements to instruction hierarchy in frontier LLMs, and equipping the Responses API with a computer environment so the model can take action on a real machine. Both are interesting on their own. In the context of the same week's news, both also describe what a Pentagon-contracted assistant would need to do.