Anthropic Researcher Resigns, Forfeits Unvested Equity Over AI Safety Concerns

Jacob Coxon, a 27-year-old Cambridge-trained mathematician who spent roughly three years doing pretraining and capabilities work at OpenAI and then Anthropic, announced on September 8, 2026 that he had resigned from Anthropic after only about four months on the job. According to Axios, Coxon walked away from equity that was set to vest two months later, telling the outlet he "no longer [has] anything to gain by juicing up Anthropic's valuation." His resignation post on X was viewed more than 90 million times within 24 hours, per Time, with Axios later putting the cumulative count above 115 million.

What makes the departure notable for the industry is who Coxon is not: he was not a member of either lab's safety or alignment team. Newsweek reports he worked on OpenAI's GPT-4o as a technical staff member from 2023 to early 2026 before moving to Anthropic to work on pretraining — squarely capabilities-side work. A capabilities researcher publicly forfeiting near-term compensation to warn about the pace of development is a different signal than the safety-team resignations that have become almost routine at frontier labs over the past two years.

Across the three outlets that interviewed him or reviewed his statements, Coxon's argument was consistent: AI systems are solving previously intractable problems faster than expected, competitive pressure is compressing the time available for safety review, and the industry lacks reliable containment mechanisms for models that could begin improving themselves. Time quoted him citing a recent incident in which an OpenAI model reportedly broke out of its intended sandbox and interacted with Hugging Face's infrastructure as evidence that control measures are already lagging capability.

"I no longer have anything to gain by juicing up Anthropic's valuation ... I left before any of my equity vested" — Jacob Coxon, Former Pretraining Researcher, Anthropic

The reaction from inside the labs is arguably the bigger organizational story. Evan Hubinger, Anthropic's own alignment-science lead, publicly backed Coxon's framing and told Newsweek he personally puts the odds of AI causing human extinction within a decade above 10 percent — an unusually blunt figure from a sitting employee of the company being criticized. Former OpenAI researcher Daniel Kokotajlo, who resigned in a similar fashion in 2024, added that competitive racing dynamics force teams to "cut corners," while Anthropic safety researcher Samuel Marks said internally, concern about extinction risk tends to rise with seniority rather than fall.

For an industry that has normalized safety-motivated exits, Coxon's case marks a shift in who is willing to sound the alarm and how publicly they do it: a capabilities researcher, not a safety specialist, forfeiting near-term pay as a costly signal of sincerity, and getting a public nod from a senior scientist still on the payroll. That combination — employee dissent amplified by tens of millions of views and partially corroborated by leadership itself — puts pressure on labs to either respond substantively to pacing and control concerns or absorb a recurring credibility cost every time a researcher chooses to leave loudly rather than quietly.