Anthropic researcher puts AI’s odds of "killing all humans" above 10%

Published Sep 9, 2026, 11:24 AM EDT Patrick O'Rourke is XDA's News Editor and Entertainment Segment Lead. Previously, he was Pocket-lint's Editor-in-Chief, the Editor-in-Chief of Canadian tech publication MobileSyrup, and earlier in his career, he worked as the technology editor at the Financial Post and Postmedia. He's based in Toronto. Over the past 15 years, he's written thousands of articles. Patrick has also interviewed dozens of tech industry executives and covered GDC, E3, Gamescom, WWDC, Apple keynotes, Samsung Unpacked events and more. Patrick has a BA in journalism from Toronto Metropolitan University. Sign in to your XDA account Summary Anthropic insider warns AI could "kill all humans," with a less than 10% chance within the next decade. Anthropic and OpenAI are racing toward self-improving superintelligence, "gambling with our lives." Current models are low risk now, but recursive self-improvement and breaches are warning shots. Welcome to The Terminator IRL. According to an Anthropic safety researcher, there's a greater than 10% chance that AI could "kill all humans." This dystopian statement comes after an employee quit at Anthropic and said that AI giants are "gambling with our lives." These concerning comments come as the AI race heats up, and both Anthropic and OpenAI continue to raise massive amounts of money while barreling towards inevitable IPOs. According to Jacob Coxon, a researcher at Anthropic who recently left the company, neither Anthropic nor OpenAI is acting responsibly in their ongoing AI development (via CNBC). In a post on Twitter, Coxon said, "They are racing straight to self-improving superintelligence and gambling with our lives." "Do not underestimate the power of this technology. These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources. We have all witnessed the progress in each of these domains, and progress is not slowing," wrote Coxon. He went on to say that "people building AI earnestly believe that it could kill us all by the end of the decade." Those odds aren't that bad... right? Superintelligence could come from recursive self-improvement This is where Evan Hubinger, an alignment science lead who still works at Anthropic, comes in with their own perspective in a recent X post. According to Hubinger, Coxon's assessment is correct, and he does "earnestly believe AI could kill all humans!" However, he also mentions that the risk from current AI models is "low." Cool, that makes me feel better. "What I am worried about is superintelligence arising from recursive self-improvement, as we have said is happening faster than we thought," wrote Hubinger. In a recent blog post, Anthropic outlined that "full recursive self-improvement" could result in "humans losing control over AI systems." We already have examples of this happening to some extent. For example, an OpenAI model breached Hugging Face back in July. In an X post, Coxon calls the Hugging Face incident a "warning shot." In other AI news, Nvidia recently acquired Hugging Face for $12.93 billion. OpenAI also recently launched GPT-6 Astra and labeled it the start of its "AGI era."