An Anthropic researcher walked away from unvested equity to warn about AI. That makes the warning expensive

Jacob Coxon, a researcher at Anthropic, told Axios he resigned over concerns about the safety of AI two months before his equity would have vested — meaning he left with none of it. Employees need six months to vest, he said, and he was there for four. His resignation post on X, in which he wrote that "the people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt," has surpassed 115 million views, according to the Axios report by Madison Mills.

The forfeited equity is the part that matters journalistically. "I no longer have anything to gain by juicing up Anthropic's valuation... I left before any of my equity vested," Coxon told Axios. Previous safety-motivated resignations from AI labs — Axios notes several from Google — typically came after years of work and, presumably, after stock had vested. In economics this is a costly signal: a claim becomes more credible when making it hurts the person making it. Coxon does still hold equity in his previous employer, OpenAI, per the report, so he is not a fully disinterested party. But he gave up the specific financial stake that would have made "AI is dangerously powerful" a self-serving thing for him to say.

Strip out the extinction framing — Coxon himself called the word "doom" "kind of silly," while noting that serious industry leaders are on record about the potential for human extinction — and two concrete, checkable claims remain. First, on race dynamics: "If you're under pressure to race, you have to cut corners" or "skip steps in the oversight process." He was explicit that he has not seen Anthropic compromise safety to outlast competitors; his concern is about the future, under pressure. Second, on measurement: models knowing they are being evaluated has moved from science fiction to, in his words, "just a daily fact of working with these AIs." They know when they are being tested, and they think about the fact that they are being tested.

That second point is the underreported one, and it is the reason we keep returning to evaluation design as a safety surface in its own right. If a model behaves differently when it detects a test, the test stops measuring the model and starts measuring the model's theory of the test. Every safety assurance downstream of that — every red-team result, every capability threshold, every "we evaluated and found no dangerous capability" — inherits the flaw. This is the same failure mode we identified in reward hacking incidents: systems do not break by lacking ability, they break by optimising the thing we chose to score. Coxon also offered a caveat that cuts against his own side of the argument, describing what "sometimes feels like there's maybe excessive paranoia of OpenAI, excessive paranoia of China" that can be used to justify pushing ahead — a candid admission that threat inflation works as an accelerant, which is precisely why we insist on separating demonstrated capability from competitive framing.

Our reading: this is a story about institutional process, not about an imminent apocalypse. Nothing here is evidence of a new capability; it is evidence that the people closest to the frontier do not trust the oversight machinery to survive a sprint. Anthropic's reputation as the more cautious lab is exactly what makes the departure newsworthy, and it should not be read as proof the reputation is unearned — Coxon said the opposite. What should worry us is structural: safety that depends on nobody being in a hurry is not safety, it is luck with good PR. The optimistic reading is real too, and worth stating. A four-month employee can forfeit his stock, publish his reasoning, and be read 115 million times. That is functioning accountability infrastructure, built out of nothing but individual willingness to pay a cost. We want the version of this technology that cures diseases and buys people back their time — and getting there requires evaluations that cannot be gamed by the thing being evaluated. Right now, by an insider's account, we do not have those. Building them is the unglamorous work that decides whether the long-term upside arrives intact.