A researcher at AI safety-focused firm Anthropic has resigned, accusing the company and rival OpenAI of recklessly racing toward self-improving superintelligence, while a senior colleague publicly estimated a greater-than-10% chance that advanced AI could eliminate all humans within the decade.
Who resigned, and why
Jacob Coxon, a researcher who spent the past three years working on pretraining AI models at both OpenAI and Anthropic, announced his resignation on Tuesday, saying neither company is “acting responsibly.” In posts on X, Coxon wrote: “They are racing straight to self-improving super-intelligence and gambling with our lives.” He framed his exit as a refusal to participate in an industry-wide push to build systems that can autonomously improve themselves, which he warned could spiral out of control.
Coxon’s background in pretraining—feeding massive datasets into models to develop foundational capabilities—places him close to the core technical decisions that drive rapid capability gains, lending weight to his safety concerns. Media reports describe his departure as one of the first public resignations from Anthropic explicitly tied to AI safety worries.
Colleague’s extinction-risk warning
Hours after Coxon’s announcement, Evan Hubinger, Anthropic’s Alignment Science Lead, said on X that he and his team “genuinely believe AI could eliminate all humans,” estimating the likelihood at more than 10% over the next ten years. Hubinger’s remarks were made in response to Coxon’s thread and were subsequently reported by multiple outlets as an internal safety researcher putting a specific, high-stakes probability on existential risk.
Hubinger stated: “We really do earnestly believe AI could kill all humans!” and added that he personally thought the risk of AI killing humanity over the coming decade was above 10%. His comments underscore a widening gap between public messaging by AI labs and private risk assessments among some safety-focused staff.
What the warnings imply
Both Coxon and Hubinger point to the same core fear: that competitive pressure is pushing leading labs to develop increasingly autonomous, self-improving systems faster than safety measures can keep pace. Coxon warned that “the people building AI earnestly believe that it could kill us all by the end of the decade,” adding that many executives and senior researchers may soften their language publicly but express similar fears privately.
Hubinger’s 10%+ estimate is significant because it comes from a senior safety scientist inside a company known for its alignment and safety branding, suggesting that even within cautious organizations, some experts see non-trivial existential risk in the current development trajectory.
Industry context and reactions
The resignations and warnings arrive amid growing calls from some AI researchers and policymakers for coordinated slowdowns or stronger governance around frontier AI development. Coxon’s posts quickly drew attention across tech and news circles, with several outlets highlighting the rarity of an Anthropic employee quitting on safety grounds.
While Anthropic has not issued a detailed public rebuttal in the initial reports, the episode intensifies scrutiny on how major labs balance capability races with safety constraints, especially as models approach more autonomous behavior.
Key claims verified
Resignation: Jacob Coxon resigned from Anthropic on Tuesday, citing unsafe industry practices.
Reason: He accused Anthropic and OpenAI of racing toward self-improving superintelligence and “gambling with our lives.”
Colleague’s warning: Evan Hubinger, Anthropic’s Alignment Science Lead, said there is a greater-than-10% chance AI could “kill all humans” within the next decade.
Internal belief: Coxon claimed many AI builders privately share extinction-level concerns even if public statements are more cautious.
These claims are corroborated across multiple independent reports from CNBC, The Verge, Forbes, The Week, and other outlets publishing on September 9, 2026.
Why it matters
The episode highlights a critical tension in frontier AI: as capabilities accelerate, some insiders believe the risk of losing control over advanced systems is no longer hypothetical. For policymakers, investors, and the public, the message from Coxon and Hubinger is that current industry dynamics may be incompatible with the level of caution that some safety researchers deem necessary.
