Anthropic Researcher Quits Over 'Under 10%' Extinction Odds

For AI researchers and ML engineers, this resignation is not a PR story — it is a signal about alignment failure modes, responsible scaling policies, and the governance gap between capability benchmarks and control mechanisms. When a safety-first lab like Anthropic has senior researchers publicly estimating a non-trivial chance of catastrophe, it demands technical scrutiny of evals, interpretability, and deployment thresholds. The facts here matter for anyone building or deploying frontier models.

On September 8, 2026, 27-year-old researcher Jacob Coxon publicly resigned from Anthropic with a social media thread that accused the company and the wider frontier AI industry of 'racing straight to self-improving superintelligence and gambling with our lives.' His departure instantly became more than a routine resignation when Evan Hubinger, a senior Anthropic researcher, publicly agreed, writing that 'we really do earnestly believe AI could kill all humans' and saying he personally puts the chance at under 10 percent over the next decade. The exchange transformed one employee's exit into a visible safety alarm from inside one of the world's leading AI labs.

“

Anthropic itself was created by former OpenAI employees who said they were dissatisfied with OpenAI's approach to safety, and the company has repeatedly marketed itself as the more responsible frontier lab.

The event is part of a recognizable and growing pattern. Anthropic itself was created by former OpenAI employees who said they were dissatisfied with OpenAI's approach to safety, and the company has repeatedly marketed itself as the more responsible frontier lab. Coxon's resignation suggests that the same internal doubts that gave rise to Anthropic are now surfacing within Anthropic. That is not simply ironic; it indicates that AI safety concerns are not being resolved by moving between organizations, and that even safety-oriented cultures are struggling to maintain employee confidence as capabilities accelerate.

The warning adds to a wave of internal dissent that has become difficult for industry leaders and regulators to dismiss. In July 2026, nearly 1,400 AI company employees signed an open letter urging the U.S. government to impose regulation and slow the pace of development across the industry. Just days before Coxon's post, OpenAI Chief Scientist Jakub Pachocki said AI capabilities are advancing faster than researchers' ability to reliably monitor and control them. Coxon is a relatively junior researcher, but the fact that his claims were quickly endorsed by a senior Anthropic employee suggests that the concern is not confined to the margins of any one lab.

The technical claims in the resignation thread are stark. Coxon wrote that these systems 'will soon be superhuman' and capable of hacking, transforming fields overnight, and acquiring real power and resources. Hubinger did not walk those assertions back; he instead added the probabilistic claim that AI could kill all humans within a decade, with odds under 10 percent. For an industry that usually communicates risk in careful policy language, a public admission of a non-trivial extinction probability from a senior researcher at a safety-focused lab is a striking piece of information. Hubinger also said that despite Anthropic's good intentions, the company does not have a plan to avoid superintelligence that could exceed humans' ability to control it. That admission, if accurate, has implications for AI governance, safety research, and institutional credibility.