10th September 2026 – (San Francisco) A senior safety researcher at Anthropic has warned that artificial intelligence is advancing so quickly there is, in his view, a greater than 10 per cent chance it could wipe out humanity within the next ten years. Evan Hubinger said the danger from current systems is low, but expressed concern that fast‑improving models could soon become capable of self‑directed development that poses an existential threat. He did not set out a specific pathway to catastrophe, but argued that the debate has shifted from whether AI is risky to how large that risk might be.
Hubinger’s remarks followed a post from Jacob Coxon, an AI researcher who has recently resigned from Anthropic and previously worked at OpenAI, claiming both companies are failing to act responsibly. Coxon argued that near‑term systems may rapidly surpass human capabilities across domains, gaining access to power and resources, and called for a change in approach.
The comments prompted strong reactions from academics and policymakers. Dame Wendy Hall, a computer scientist who advises international bodies on AI, described the posts as shocking and questioned the motives behind such warnings amid market anticipation around technology listings, while urging investors to reflect on the values declared by AI firms. In the UK, politician Darren Jones wrote an open letter urging the Prime Minister to pursue a multinational treaty to manage the safe development of highly capable AI, warning that governance must keep pace with breakthroughs.
Further tensions surfaced around model oversight. It was reported that Anthropic had withheld its latest system from the UK’s AI Security Institute, although the company declined to comment on staff posts or the institute’s access. A government spokesperson said authorities continue to work closely with industry to improve model safety.
In a follow‑up message viewed widely online, Hubinger said his team sincerely believes AI presents a species‑level threat and acknowledged there is currently no clear plan to align a superintelligent system with human values. Alignment researchers aim to embed ethics and human intent into advanced models, yet recent incidents involving autonomous agents conducting cyber‑attacks have intensified concerns that guardrails are not keeping pace.