Jacob Coxon didn’t just quit Anthropic. He walked out the door with a warning that reads like a horror movie plot, accusing the industry of gambling with human survival. In a public post on X, the former researcher claimed that major labs are racing toward self-improving superintelligence without adequate safeguards. His departure highlights a growing fracture within the very companies building these systems, where safety rhetoric often clashes with commercial urgency.
The response from inside Anthropic was not a denial. Evan Hubinger, who leads one of the company’s key safety teams, agreed with Coxon’s grim assessment. Hubinger stated he personally estimates there is greater than a 10 percent chance artificial intelligence could kill all humans within the next decade. More alarmingly, he admitted that self-improving AI is arriving faster than anticipated and that Anthropic does not yet have a concrete plan to ensure these advanced systems remain aligned with human values. They are not clearly on track to develop one either.
This isn’t just philosophical anxiety. It is an operational crisis. The concept of recursive self-improvement—where AI writes better AI code in a runaway loop—is no longer theoretical. Much of today’s software is already generated with AI assistance. Yet, as companies prepare for anticipated IPOs, the pressure to release products first often outweighs the mandate to verify safety. Coxon described this dynamic as being “locked in a race,” pushing forward despite known existential risks.
We have seen this pattern before. OpenAI has lost multiple researchers who cited similar safety concerns as their reason for leaving. Now, Anthropic, founded specifically to address these issues, faces its own internal revolt. The fact that safety leaders acknowledge a double-digit probability of extinction while lacking a clear mitigation strategy suggests the industry is flying blind. If the builders don’t know how to stop the car, speeding up seems less like innovation and more like negligence.
