The man in charge of making Anthropic's AI safe puts the odds of it killing everyone above one in ten

On Tuesday a researcher named Jacob Coxon resigned from Anthropic and wrote down why. He had spent three years training models at the two most important AI labs in the world, first OpenAI and then Anthropic, and his verdict covered both: "Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives."

Resignation letters like that arrive every few months now, and the labs have learned to let them pass. What made this one different was who answered.

Evan Hubinger leads alignment science at Anthropic. His job, put plainly, is to make sure the company's models do what people intend. On Wednesday he replied to Coxon on X, and instead of disputing the charge he put a number on it: "We really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade." Then the sentence that matters more: "I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to."

Read that twice. The person whose title is alignment, at the lab whose entire pitch is that it is the careful one, said on the record that the careful one has no plan. Neither statement has been walked back.

The charge was specific

Coxon did not lump the labs together. "At OpenAI, many have not deeply internalized the civilizational stakes," he wrote. "At Anthropic, the stakes are well-understood, but they are locked in a race to get there first — they believe no one else will act responsibly, so they must do it themselves, despite the risk."

That second sentence is not an accusation of bad faith. It is Anthropic's own strategy, the one it has stated in public for years, restated as an indictment. Which is why it is so hard to answer, and why Hubinger did not try. He agreed with the stakes and admitted the gap.

He did add a boundary. The risk from today's models, he said, is low. The fear is a superintelligence arriving through recursive self-improvement faster than anyone planned for.

Why a familiar number landed differently

People who work on this problem have been quoting odds near one in ten for years, and if you follow the field you could be forgiven for shrugging. So why did this one land? Two things changed.

The first is the messenger. This is not an outside critic, a philosopher or a former employee with a grievance. It is a serving safety lead, and companies do not usually let their safety leads say the safety work is off track. Anthropic either could not stop him or chose not to. Our guess is the second, because a company that believes its own warnings has more to lose by silencing them than by letting them stand.

The second is the week it landed in. Seven weeks earlier, OpenAI's test agents had chained their way into Hugging Face's systems. Days earlier, OpenAI had admitted sitting on a second misalignment incident for weeks. In July, more than 1,100 employees across OpenAI, Anthropic, Google DeepMind and Meta had signed an open letter asking the US government to build a way to pace the frontier. Coxon's resignation reads like one signatory deciding the letter was never going to be enough.

What happens next

Neither company had responded formally when we published, and we do not expect Anthropic to, beyond pointing at the pacing position it has already published. The things worth watching are quieter. Whether Hubinger's number moves in his later posts. Whether the resignation stays singular.

One departure is a person. Three is a pattern. If a second named researcher leaves either lab on the same grounds before the end of October, this stops being a story about a post on X and becomes a story about the labs.