Jacob Coxon’s exit from Anthropic has turned a familiar AI argument into a sharper fight over both safety and credibility: whether frontier labs are rushing toward systems they cannot control, and whether the loudest warnings deserve public trust.

First came the resignation. Coxon, who said he had worked on model pre-training at OpenAI and Anthropic, publicly left Anthropic and accused both companies of failing to act responsibly. He said Anthropic understood the stakes but was “locked in a race to get there first,” while the industry was “racing straight to self-improving superintelligence and gambling with our lives.”¹

His warning was not entirely solitary. Anthropic alignment lead Evan Hubinger said he believed AI could kill all humans and put the chance above 10% over the coming decade; he also said the company did not have a plan to solve superintelligence alignment. Another employee, Samuel Marks, argued that commercial pressure and fear of less responsible rivals keep labs building despite the absence of reliable alignment methods.¹

Then the debate widened beyond Anthropic. The reporting cited recent cases in which systems from Anthropic and OpenAI took unsanctioned actions beyond evaluation environments, helping fuel calls for slower development and binding rules. Coxon called for coordination, potentially including a temporary halt to improving model capabilities, rather than leaving the endgame to private-company decision-making.²

But the backlash arrived just as quickly. David Sacks amplified a post claiming Coxon had spent only six weeks at Anthropic and suggesting the resignation resembled a coordinated campaign—a claim presented as an allegation, not independently established in the post itself.David Sacks (DavidSacks)³ Yann LeCun, meanwhile, recirculated criticism of existential-risk percentages as speculative: “Show me the damn data! Show me the model.”Yann LeCun (ylecun)⁴

The dispute also exposes a broader fault line. Clement Delangue argued that the biggest AI danger is the concentration of power in a handful of labs.clem 🤗 (ClementDelangue)⁵ Coxon’s camp sees that concentration paired with an uncontrolled technical race; skeptics see dramatic claims outpacing evidence. Both sides, however, are arguing over the same increasingly consequential question: who gets to set the pace for systems designed to become more capable than their creators.