Anthropic’s Evan Hubinger, who leads one of the company’s AI safety teams, says there is a greater-than-one-in-10 chance that artificial intelligence could kill all humans within the next decade.
Hubinger made the estimate after Anthropic researcher Jacob Coxon announced his resignation, accusing AI companies of racing towards self-improving superintelligence despite believing the technology could become uncontrollable. The exchange was reported by The Verge, which reviewed the researchers’ posts on X.
Anthropic AI safety warning follows researcher’s resignation
Coxon, who previously trained AI systems at OpenAI, said he was leaving because the companies were “racing straight to self-improving superintelligence and gambling with our lives”. He argued that AI labs were locked in a race to develop advanced systems first, even as their own staff believed the systems could pose an existential risk.
Hubinger agreed with the substance of Coxon’s warning. “We really do earnestly believe AI could kill all humans,” he wrote, adding that he personally estimated the chance at more than 10 per cent within the next decade.
He also said Anthropic did not yet have a plan for ensuring advanced AI remained safe and aligned with human values, and was not clearly on track to develop one. The comments are unusually direct for a company whose public identity is built around reducing the risks of advanced AI. The Anthropic AI safety warning is a personal assessment from a senior researcher, not a company forecast.
What the AI safety warning does — and does not — claim
The researchers were discussing a future risk, not reporting that an AI system had killed people or escaped operational control. Their concern centres on self-improving systems: models that could help develop more capable successors in a cycle that becomes difficult for people to monitor or stop.
The BBC’s reporting separately describes Anthropic as a company that has positioned itself as more safety-oriented than its rivals, while also pursuing commercial AI products such as Claude. Its report covered another Anthropic safety researcher’s resignation earlier this year, but does not independently verify Hubinger’s probability estimate.
For companies using frontier AI, the warning puts governance, monitoring and limits on autonomous systems alongside capability and cost. It does not provide a timeline for a catastrophe, nor does it establish that the estimate is a consensus view across the industry.
The AI safety warning is based on the researchers’ personal assessments. Anthropic had not published a technical forecast assigning the same probability in the sources reviewed for this article.
Who estimated that AI could kill all humans?
Evan Hubinger, who leads one of Anthropic’s AI safety teams, said he personally estimated the chance at more than 10% within the next decade.
Why did Jacob Coxon leave Anthropic?
Coxon said AI companies were racing towards self-improving superintelligence and gambling with people’s lives despite the risks.
Is the 10% figure an official Anthropic forecast?
No. It is Hubinger’s personal assessment, and the sources reviewed do not show Anthropic publishing the same probability as a company forecast.
