Anthropic researcher resigns, warns AI labs are 'gambling with our lives'

Jacob Coxon, a pre-training researcher, resigned from Anthropic on September 9, 2026, warning that the industry's pursuit of superintelligence poses an existential threat. The departure marks a high-profile break from one of the world's leading AI safety labs.

In a viral thread on X, Coxon accused both Anthropic and OpenAI—where he also spent time during his three years of pre-training research—of recklessly racing toward self-improving superintelligence. He claimed the labs are "gambling with our lives" and called for immediate pacing agreements between U.S. laboratories. To prevent a global race toward uncontrollable AI, Coxon advocated for potential temporary bans on capability advancements.

The Internal Rift

Coxon's resignation comes as a wave of new startups, including Ricursive Intelligence, Recursive Superintelligence, and Discovery Loop, aggressively pursue recursive self-improvement. This competitive environment has created a stark divide between corporate branding and technical reality. Coxon alleged that while executives and senior researchers privately express fear of human extinction, they publicly couch their phrasing to appear sensible.

This internal tension was underscored by admissions from Evan Hubinger, Anthropic's Alignment Science Lead. Hubinger admitted that the company currently lacks a plan to solve the alignment problem for superintelligence. Hubinger further stated his personal belief that there is a greater than 10% chance that AI could kill all humans within the next decade.

Industry Implications

The event highlights a critical rift at "safety-first" labs, where researchers believe the competitive dynamic forces them to ignore existential risks in favor of speed. The admission that a leading safety lab has no concrete plan for superintelligence alignment suggests that the technical safeguards currently in place may be insufficient for the systems being developed.

This instability has already triggered legislative responses. In the U.S., lawmakers including Bernie Sanders and Greg Casar have discussed the "Ban Artificial Superintelligence Act" as of September 2026, reflecting a growing political appetite to regulate or prohibit the development of autonomous superintelligent systems.

What's Next

Industry observers are now watching to see if other researchers at major labs will follow Coxon's lead or if the proposed pacing agreements will gain traction among U.S. AI firms. While the push for legislative bans grows, it remains unconfirmed whether these efforts will result in enforceable law or if the momentum of recursive self-improvement will outpace government intervention.