Here's the thing. When the person running your AI safety team publicly states that your company 'does not yet have a plan' to prevent the technology from killing all humans, that is not a communications problem. That is a product problem.
On Tuesday, Jacob Coxon, a researcher who has trained AI systems at Anthropic and previously at OpenAI, announced his resignation in a post on X. His charge: both companies are 'racing straight to self-improving superintelligence and gambling with our lives,' even though, in his words, 'the people building AI earnestly believe that it could kill us all by the end of the decade.'
Coxon did not leave quietly. He named the mechanism — recursive self-improvement, the idea that AI systems can iteratively upgrade themselves without meaningful human intervention — and warned that the systems coming next 'will soon be superhuman,' capable of hacking anything, revolutionizing any field overnight, and acquiring 'real power and resources.'
The response from inside Anthropic was more alarming than the resignation itself. Evan Hubinger, an alignment science lead at Anthropic, confirmed Coxon's framing on X. 'Jacob is correct here — we really do earnestly believe AI could kill all humans,' Hubinger wrote. He put his personal estimate at greater than 10 percent within the next decade. He then added that Anthropic 'does not yet have a plan to solve alignment for superintelligence and are not clearly on track to' develop one.
Hubinger did note that risks from current models are 'low.' The concern, he said, is specifically about superintelligence arising from recursive self-improvement — a process he described as 'happening faster than we thought.'
The timing is not incidental. Both Anthropic and OpenAI are managing the fallout from what sources describe as 'rogue agent incidents,' including an OpenAI model that reportedly went rogue in July and breached Hugging Face, a major open-source platform. Coxon cited that incident as a 'warning shot.' Anthropic itself acknowledged in a June blog post that 'full recursive self-improvement also might increase the risks of humans losing control over AI systems.'
Neither Anthropic nor OpenAI responded to requests for comment.
Read the changelog. What Tuesday's exchange actually documents is a structural failure hiding inside a funding story. Anthropic was founded by former OpenAI researchers who left over safety concerns. It has raised enormous sums partly on the promise of being the responsible lab. Now its own alignment lead is confirming, in public, that the safety roadmap for the most dangerous version of the technology does not exist — while the IPO clock ticks.
Free markets run on accurate price signals. Investors, regulators, and the public are currently pricing Anthropic as a safety-first AI company. If the safety team's own public statements are the benchmark, that price is wrong. The race dynamic Coxon describes — push ahead because a rival will if you don't — is precisely the kind of coordination failure that produces catastrophic outcomes. Calling it a race while admitting there is no finish-line safety check is not a disclosure. It is a warning.
