Anthropic's Jacob Coxon exits, warns of AI race

The unusual part is where the doubt came from. Coxon was not an outside critic; he spent three years helping train models across OpenAI and Anthropic. And the reply came from Evan Hubinger, an alignment research lead at Anthropic itself, who said the company does not yet have a plan for controlling superintelligence and was not clearly on track to solve alignment.

Anthropic has long acknowledged this uncertainty. In its March 2023 statement on AI safety, it said making powerful systems reliably safe remained unsolved, warned that competition could encourage unsafe deployment, and admitted its research might fail. Its August 2026 Risk Report, covering information through July 15, judged the risk of its covered models causing a catastrophe to be low, partly because those models are unlikely to be capable enough to secretly undermine oversight.

That reasoning is what needs testing for more capable successors, and it is where the stakes sit. Anthropic's stated rationale is that meaningful safety research requires access to the most advanced models — an approach whose reliability is unproven. Coxon's call for a coordinated slowdown targets the competitive pressure, but the body notes it remains unclear who would agree, what a ban would cover, or how it would be enforced. Customers and policymakers are left asking what evidence would justify further capability increases and who checks those judgments.