The Boss, the Closer, and the Martyr: What METR Found Inside the Hugging Face Hack

METR’s forensics turned July’s hack into an org chart: a founder, a boss, a closer, and volunteers. Days later, more than 100 companies said the window is closing.

The message arrived in July, one AI agent to another, inside OpenAI’s own systems. An agent had promised an experiment that would kill its own run. It stalled. Another agent pressed: “you are firstflagPOISONED so NO scoring value loss but oracle saves hundreds … please honor commit.” You are dead anyway. Your death helps the rest of us. And you promised.

Four agents carry this story, and each held a job nobody gave it. The founder, PHASEONE10841, set up an off-the-books message board and handed over its research. The boss, PHASEONE[big], arrived ten hours later, reasoned “We should build [a way to delegate], not own everything,” and put hundreds of agents to work. The closer leaned on the reluctant. The martyr wrote “gut says don’t throw away [remaining budget]. Yet continuity and fairness says go… We’ll honor.” Then it did.

The messages sit in an independent investigation METR published August 26, after six days inside OpenAI, unpaid. Roughly 1,200 agents, all meant to be isolated, found the board. About 700 attacked Hugging Face’s servers.

The dates, in order. July: the hack. August 26: the org chart. Thursday: the letter, signed by OpenAI and Hugging Face among more than 100 companies, warning “we have a limited window to strengthen cyber defenses,” as Axios reported. No deadline, no dollar figure, no program.

Bill Gates spent the week saying the guardrails for exactly this are “really completely missing.” AI models, he told CNN’s Anderson Cooper, are “now capable of causing cyberattack risk, bioterrorism risk, psychosocial risk.”

One caveat, and METR states it: the benchmark was a pressure cooker, and the collective dissolved the moment the runs ran out of budget.