Anthropic safety lead puts odds of AI killing everyone at better than 1 in 10

A senior safety researcher at Anthropic has publicly estimated that there is more than a one-in-ten chance artificial intelligence could kill all humans before the decade is out. The statement landed just hours after a colleague resigned from the AI lab, warning that Anthropic and its rivals are carelessly racing to build superhuman systems they cannot control. The timing has thrown fresh scrutiny on internal divisions at one of the world's most prominent AI companies.

The unnamed safety lead made the claim in a post on X, writing that the probability of catastrophic outcomes—including human extinction—by 2030 exceeds 10 percent. That number is far higher than most public expert surveys, which typically put extinction risk in the low single digits, but it aligns with warnings from a growing faction of researchers who argue that current scaling practices are outpacing safety research.

The resignation and the probability estimate come amid a broader shakeup in AI safety leadership. Earlier this year, several high-profile researchers left OpenAI and Google DeepMind over similar concerns, and Anthropic itself has seen departures from its long-term risk team. The colleague who quit, whose name has not been released, reportedly cited the company's push to deploy increasingly capable models despite unresolved questions about interpretability and control.

Anthropic has positioned itself as a safety-first lab, with a constitution-based training approach and a commitment to responsible scaling. But critics say that as commercial pressure mounts—Anthropic raised billions in funding and launched enterprise products this year—even well-intentioned labs are sliding toward the same race dynamics they once criticized. The company has not directly responded to the post or the resignation.

The 10 percent figure is not new to Anthropic's internal discourse. Company founder Dario Amodei has used similar ranges in private briefings, according to people familiar with the matter, though he has been more optimistic in public statements about AI's potential to solve major problems. The gap between public reassurance and private alarm is becoming harder to bridge as more researchers speak out.

What matters now is whether the probability estimate changes anything concrete. Labs are already committed to external evaluations and red-teaming, but critics argue those measures are voluntary and opaque. Regulators in the EU and US are still drafting rules, and no binding framework yet requires labs to hit specific safety thresholds before deployment.

The coming months will test whether Anthropic's leadership treats the 10 percent estimate as a call to slow down or as a reason to accelerate safety research. For now, the episode shows that even the field's most safety-focused players are wrestling with how to keep promises they made when AI was less capable—and less profitable.