Anthropic researcher Jacob Coxon quits, warns AI could kill humans

Jacob Coxon, a researcher who spent years working on AI safety at Anthropic, has resigned and gone public with a warning that the systems he helped build could eventually kill people. In an interview with Politico, Coxon said he could no longer in good conscience stay at the company, which he argues is favoring capability over caution. His exit lands as one of the more pointed insider critiques yet from a lab that has built its public identity around safety.

Coxon's argument is not that a single disaster has already happened. He points instead to a chain of ordinary product decisions that, taken together, shrink the margin for error — and to a technical gap he says nobody has closed. "We are building machines that could outsmart us, and we don't have a proven way to keep them aligned," he told Politico. He singled out loss-of-control scenarios, in which a model behaves in ways its developers did not intend and cannot reliably undo, and said benchmarks, red-teaming and alignment testing are not enough to rule those out.

The timing puts Anthropic in an awkward spot. Founded in 2021 with a stated mission of prioritizing safety, the company has positioned itself as the more careful alternative to OpenAI and Google DeepMind. It recently shipped its Claude 4 family, which pushed further on reasoning and autonomous action — the exact capabilities Coxon and other safety researchers have flagged as risky when released before the underlying control problem is solved. A company that sells itself on restraint now has a former employee saying commercial pressure won out.

Coxon's departure is one entry in a longer list. Similar public resignations have come out of OpenAI and Google DeepMind in recent months, and in 2024 a group of former employees from OpenAI and Anthropic signed an open letter asking for more transparency and stronger whistleblower protections, saying internal safety concerns were being suppressed. That pattern is what gives the exit weight beyond one person's opinion: the people leaving are the ones who were paid to think about failure modes.

Reaction inside the research community has split. Some praised Coxon for saying in public what they hear in private; others called the warning alarmist and pointed to real progress Anthropic has made in interpretability, the subfield that tries to see inside a model's internal state rather than judge it only by its outputs. Both things can be true at once, which is part of why the dispute has no clean resolution.

Coxon has not said where he will go next, only that he intends to keep working on safety outside a corporate lab, through independent research or advocacy. The open question is what his exit changes, if anything. Labs are still racing to train larger models on larger clusters, and no resignation has yet moved a release date — so the test ahead is whether Coxon's warning, and the ones before it, shift how frontier labs document and gate the capabilities they ship.