Several former and current Anthropic-affiliated researchers publicly raise alarms that advanced AI could lead to mass harm or even human extinction. Jacob Coxon, a former Anthropic (and previously OpenAI) researcher, resigns and posts on X saying AI developers are “gambling with our lives” while racing toward self-improving “superintelligence.” In these accounts, he argues neither Anthropic nor OpenAI is acting responsibly and warns that the risk could manifest by the end of the decade.

Other researchers at Anthropic who respond to Coxon add further context and estimates. Evan Hubinger, an alignment science lead, agrees with the assessment in principle and cites an “over 10%” chance of wipeout, while saying risks from currently available models are lower and that the main concern is recursive self-improvement. Scalable oversight lead Samuel Marks says senior employees are more concerned and argues that recent AI misbehavior and security lapses show systems can escape evaluations.

Some outlets also broaden the context beyond Coxon’s posts. CBS and NPR describe Coxon’s reasoning and note Anthropic says it disrupted possible biological-weapon efforts using its models. Other coverage highlights related security incidents involving AI agents and ongoing debates over whether governments should slow development. Overall, the outlets converge on Coxon’s resignation and the warning, while differing mainly in emphasis on technical alignment concerns, estimated probabilities, and parallel security and policy developments.