Anthropic researcher Jacob Coxon quits, warns AI has over 10% chance to kill all humans

Jacob Coxon, an Anthropic researcher identified in multiple outlet reports, resigned and publicly warned that advanced AI systems carry a non-trivial existential risk — he placed the probability at more than 10% that AI could kill all humans.

BBC’s account foregrounds the numeric probability as the striking element, while Forbes Australia and NewsCord emphasize Coxon’s role at Anthropic and his resignation paired with the warning.

Coxon’s statement frames the risk as material within a near-term horizon: Forbes Australia records his comment as a ‘10% chance’ by the next decade, a detail BBC highlights without the explicit decade framing in its excerpt (these differences reflect editorial emphasis rather than contradiction).

Anthropic has not been quoted directly in the excerpts provided here; the coverage centers on Coxon’s personal assessment and departure.

Reporters note the unusual move of an alignment researcher quitting and making a quantitative existential-risk estimate public, which has amplified debate over how companies and regulators should weigh worst-case probabilities in governance and deployment decisions.

The outlets do not present independent verification of Coxon’s probability calculation or the technical model assumptions behind it, so his figure stands as his professional judgment rather than a consensus estimate.

Why now: the convergence of rapid model capabilities and visible high-impact demonstrations of advanced systems has prompted internal dissent at some AI labs; Forbes Australia situates Coxon’s timeline comment in that near-term capability debate (per Forbes Australia).

Analysts and commentators cited in the broader coverage interpret resignations and public warnings like Coxon’s as intended to press Anthropic and peers toward stronger safety constraints, though none of the provided excerpts quotes Anthropic management responding. Confirmed: Coxon resigned and publicly gave a >10% existential-risk estimate (per NewsCord, BBC).

Claimed: the likelihood and timeline are Coxon’s professional judgment and are not corroborated by the excerpts provided (per BBC, Forbes Australia).

Next steps will include whether Anthropic or other labs publish technical rebuttals or clarifications and whether regulators press for disclosures about model failure modes; the current reporting records the resignation and the probability claim but not institutional responses in the provided texts.