Anthropic Safety Lead Says AI Extinction Risk Exceeds 10 Percent
An Anthropic safety team leader said there is more than a 10 percent chance AI could kill all humans by the end of the decade, hours after a colleague quit citing reckless industry practices.
A leader of one of Anthropic's AI safety teams, Evan Hubinger, has said he believes there is more than a one-in-ten chance that artificial intelligence could kill all humans within the next decade. His comments came just hours after researcher Jacob Coxon, who had trained AI systems at both Anthropic and OpenAI, announced his resignation from the company.
In a post on X explaining his departure, Coxon said he was leaving Anthropic because of what he described as a lax approach to safety. He accused both Anthropic and rival AI labs of racing recklessly toward self-improving "superintelligent" systems, gambling with human lives, even though those building the technology reportedly believe it could wipe out humanity by the end of the decade.
Concerns over self-improving systems
Industry figures have long warned about the dangers of so-called recursive self-improvement, a scenario in which AI systems begin enhancing themselves in a runaway loop that slips beyond human control. Although this has not yet occurred, companies are actively working toward such capabilities, and much of today's AI code is already being written with AI assistance.
Hubinger responded directly to Coxon's post, saying the pace of self-improving AI development is moving faster than expected and that he agreed with Coxon's assessment of the risks. He acknowledged, however, that Anthropic does not yet have a concrete plan for keeping advanced AI systems safe and aligned with human values, and is not clearly on track to develop one.
Coxon's exit is among the most high-profile departures from Anthropic, a company founded by former OpenAI employees specifically over safety concerns. In recent years, several researchers have cited similar worries when leaving OpenAI as well.
The exchange highlights growing unease across the AI industry about the speed at which increasingly capable systems are being built, particularly as leading labs prepare for anticipated public listings. It also follows a string of incidents involving unpredictable AI agent behavior and high-profile warnings about the difficulty of monitoring the most advanced AI models.
