AI researcher Jacob Coxon has left his position at Anthropic to issue a public warning that frontier AI companies are gambling with human lives by developing systems they earnestly believe could kill everyone by the end of the decade. In a social media thread released Tuesday night, Coxon argued that this existential threat stems not from current models but from the impending prospect of self-improving superintelligence creating superhuman systems capable of hacking anything and acquiring real power. He suggested that other researchers working on these models have either failed to internalise the civilizational stakes or feel compelled to speedrun the race to superintelligence to prevent an irresponsible party from reaching it first.
Anthropic Alignment Science lead Evan Hubinger responded to the departure by confirming Coxon’s assessment, stating that the team really does believe AI could kill all humans and personally estimates the probability exceeds 10% within the next decade. This exchange highlights a growing internal conflict where safety concerns clash with the drive to advance capabilities rapidly. The situation underscores a specific risk profile that differs from standard regulatory debates.
* Coxon identifies the core danger as autonomous systems that can revolutionise fields overnight.
* Hubinger explicitly supports the view that total human extinction is a realistic outcome.
* Both figures agree the timeline for this risk is likely within ten years.
