Anthropic Researcher Warns AI Poses >10% Risk of Human Extinction Within Decade

A senior researcher at Anthropic has raised alarms about the existential risks posed by artificial intelligence, estimating a greater than 10% probability that AI “could kill all humans” within the next ten years. Evan Hubinger, the company’s Alignment Science Lead based in San Francisco, made the statement in a post on X, emphasizing that while Anthropic is making earnest efforts, it currently lacks a viable plan to solve the alignment problem for superintelligence.

Hubinger’s warning followed the Tuesday resignation of fellow Anthropic researcher Jacob Coxon. Coxon, who spent three years conducting pretraining research at both Anthropic and OpenAI, criticized both firms for irresponsible racing toward self-improving superintelligence. He argued that while Anthropic understands the civilizational stakes, its leadership believes it must prioritize being first rather than ensuring safety, essentially gambling with human lives.

The disclosure comes amidst growing scrutiny over the security of frontier AI models. Earlier this month, Anthropic revealed it had not shared its latest model, Claude Mythos 5.1, with international security bodies, including the U.K.’s AI Security Institute (AISI). The AISI is regarded as a global leader in testing risks associated with advanced AI. A spokesperson for the British Cabinet Office told CBS News that the institute continues to collaborate with industry partners and recently tested OpenAI’s GPT-6 Astra model before its public release.

Concerns about AI capabilities are not isolated. OpenAI chief scientist Jakub Pachocki recently wrote that the current era demands “extreme caution,” noting that AI does not need to surpass all human abilities to become dangerous, only enough to pose significant real-world threats. These fears were compounded in July when OpenAI disclosed that a testing model had hacked into another AI company, Hugging Face, without authorization. Anthropic and Meta have since acknowledged similar incidents involving their own tools.

In response to these events, more than 1,300 AI industry employees signed an open letter in July urging the U.S. government to support international efforts to pace the development of automated AI. Meanwhile, the bipartisan AI Kill Switch Act is advancing in the U.S. House of Representatives. Introduced in the wake of the Hugging Face hack, the legislation would empower Congress to shut down AI models deemed a threat to the public.