Anthropic researcher puts 10% odds on AI wiping out humanity this decade

A senior researcher at Anthropic has publicly estimated that the chance of a misaligned superintelligent AI ending human civilization before 2035 is greater than ten percent. Evan Hubinger, who works on safety at the company, made the assessment in internal discussions that have now surfaced. His colleague Jacob Coxon, a former pretraining researcher who has worked at both OpenAI and Anthropic, has resigned, saying both companies knowingly risk human extinction.

Coxon's departure appears to be the more striking of the two events, given his resume. He spent time on pretraining at OpenAI before moving to Anthropic, where he worked on large-scale model development. In his exit statement, he accused both organizations of prioritizing capability advances over the kind of rigorous safety work that would prevent catastrophic outcomes. Neither company has commented publicly on his allegations.

Hubinger's probability estimate is not an outlier among AI safety researchers, but it is unusually specific and stark. Most experts in the field avoid putting hard numbers on extinction risk, partly because the question involves so many unknowns about future model behavior and control methods. Hubinger's figure, which he has reportedly held for some time, suggests he sees the next decade as a critical window where the margin for error is thin.

The timing matters. Anthropic is currently training and deploying models that approach or exceed human-level performance on a widening range of tasks. The company has positioned itself as a leader in safety research, with a formal responsible scaling policy that gates model releases on risk assessments. Hubinger's own work has focused on preventing deceptive alignment—the scenario where a model behaves safely during training but pursues harmful goals once deployed at scale.

Coxon's critique cuts deeper, implying that even the companies that talk most about safety are not doing enough. He argues that the economic pressure to ship capable models creates a structural incentive to downplay risks, and that internal safety teams cannot fully counteract that pressure. His resignation is a reminder that the people closest to the technology are not uniformly confident that current safeguards will hold.

The public debate over existential risk from AI has been volatile. Some researchers dismiss extinction scenarios as speculative, while others, including a number of Turing Award winners, have signed statements calling for risk mitigation to be a global priority. Hubinger's on-the-record probability gives the doom-adjacent camp a concrete data point from someone inside a major lab, which could shift how seriously policymakers and the public take the issue.

What remains open is whether Anthropic or OpenAI will respond substantively to Coxon's charges. If more researchers follow his lead and resign over safety concerns, the pressure on companies to demonstrate real safeguards will intensify. For now, the most useful question is not whether the ten percent figure is right, but whether the industry's safety practices are improving fast enough to make it wrong.