Evan Hubinger Backs Coxon’s AI Warning With a 10%+ Extinction Estimate

Anthropic alignment-stress-testing leader Evan Hubinger has publicly backed former colleague Jacob Coxon’s warning about the AI race. Hubinger said he personally sees a greater than 10% chance that AI could kill all humans within the next decade, and that Anthropic does not yet have a plan to solve superintelligence alignment.

Hubinger’s response came after Coxon resigned from Anthropic, where he had worked on pretraining after time at OpenAI. Coxon accused both companies of racing toward self-improving superintelligence and gambling with human lives. That is Coxon’s assessment of the companies’ direction, not evidence that such a system exists today.

Greater than 10%Chance AI kills all humans within a decade

Hubinger said Coxon was correct, while stressing that his probability estimate was his personal forecast about a future outcome.

The warning reached beyond AI circles

CNET counted more than 156 million views and more than 752,000 likes on Coxon’s X thread at the time of publication. Illinois Gov. JB Pritzker responded by calling for immediate action from the AI industry and Washington. Coxon’s warning also reached CNN and Fox News, a sign that an internal AI-safety dispute had moved into mainstream political conversation.

A risk estimate without a settled control plan

Alignment is the effort to make an AI system reliably follow intended human goals. Hubinger said Anthropic is trying its best, but is not clearly on track to solve alignment for superintelligence. His statement puts a senior member of the company’s safety work alongside Coxon’s broader concern: that competitive pressure may outrun the ability to control more capable systems.

The next test is a verifiable limit

Coxon has proposed that OpenAI and Anthropic coordinate to limit recursive self-improvement, or using AI to create better AI systems. Anthropic told WIRED it supports a lawful, verifiable way for the industry to pace releases of powerful models. Neither is an operating agreement. The concrete test for labs and policymakers is whether those broad calls become defined limits that can be checked before more capable systems are released.

The full extinction scenario remains an extreme and contested forecast, and it is unclear whether current technology can produce superintelligence. But Hubinger’s intervention changes the immediate debate: it is no longer only about a departing researcher’s warning, but about whether leading labs can specify the safety conditions they say must govern the race.