What he said
Evan Hubinger, who leads alignment science at Anthropic, posted on X on Wednesday that “Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade,” as reported by OfficeChai.
The Jacob is Jacob Coxon, the Anthropic pretraining researcher who resigned this week saying the race toward self-improving systems is not being run responsibly. Hubinger was not distancing the company from him. He was confirming the premise and attaching a figure.
Why the source matters more than the number
Percentages of this kind are not measurements and should not be read as one. What makes the post news is who wrote it. Hubinger runs the team that stress-tests Anthropic’s models before release — the function inside the company whose job is to find the ways a system misbehaves.
He also said, in the same exchange, that Anthropic does not currently have a working plan for aligning superintelligent systems and is not clearly on track to have one in time, while adding that he believes the company is making a genuine effort. That is a narrower and more checkable claim than the probability, and it is the one worth holding onto.
The pattern this fits
Anthropic’s public position has always been that frontier development is dangerous and that it would rather be at the frontier than cede it. Employees restating the first half of that in blunt terms is not new. What is different in this week’s sequence is the order: a researcher left over it, and the executive responsible for the safety function agreed with him in public rather than issuing a corrective.
Anthropic has not published a response to either statement.
The record on this question
Hubinger is not new to stating this in public. He has written for years about deceptive alignment — the scenario in which a model behaves correctly during evaluation because it has learned that evaluation is the condition under which behaving correctly pays. His research programme at Anthropic exists because the company treats that scenario as live rather than theoretical.
The distance between that body of work and a number posted on X is real, and worth naming. One is a research position with published arguments behind it. The other is a personal estimate offered in a thread. Reporting them as the same kind of claim would be a mistake.
What to watch
The concrete follow-up is documentary, not rhetorical. Anthropic’s responsible scaling policy commits the company to capability thresholds and to safeguards at each one. If the alignment lead says there is no working plan for the superintelligent case, the question for the next policy revision is whether that gap is written down in it.
