Anthropic researcher quits over AI race fears: safety work is ‘nowhere near enough’

Anthropic pretraining researcher Jacob Coxon said on X that he’s resigned, is leaving AI entirely, and thinks Anthropic and OpenAI are pushing toward more powerful systems even though the risks still aren’t resolved.

In the same thread, Coxon, 27, said he previously spent about three years at OpenAI. He wrote that the safety work at the big labs is sincere, but nowhere close to enough; that researchers can see the danger and keep building anyway because they worry a rival will get there first; and that decisions with stakes this high shouldn’t be left to private companies.

Anthropic alignment lead Evan Hubinger replied publicly on X and said Coxon was “correct.” Hubinger added that his own estimate of advanced AI killing all humans within a decade is above 10%, that Anthropic is doing its best but still doesn’t have a clear plan for aligning superintelligence, and that what scares him is not today’s chatbots but self-improving systems.

For anyone who follows AI policy or has money in the sector, this exchange is one to keep an eye on. Hubinger was giving his personal view, not an official Anthropic forecast, at least in that back-and-forth on X. Experts still disagree sharply about extinction risk. And the fight between safety and shipping could get louder as Anthropic moves closer to its IPO plans. Read the exchange on X in the broader frame too: jobs, uneven protections across languages, perspectives outside the West, and global rules being shaped by the US-China rivalry and the EU.