Anthropic’s founder warns AI could self-improve by 2028 and admits underestimating the risks.

According to Beating Monitor, Jack Clark, co-founder of Anthropic, warned during a speech at Oxford University that AI is highly likely to achieve "recursive self-improvement" by 2028 or even earlier. This means large models will be capable of autonomously upgrading themselves and generating even more powerful versions without human intervention. Clark bluntly stated that most people worldwide still deny the pace of technological self-evolution, and even Anthropic itself has severely underestimated the scale of technological advancement, leaving it unprepared for the risks of uncontrolled frontier models. Regarding the capabilities of frontier models, Clark cited the internal model Mythos, trained in April 2026, to illustrate the real boundaries of loss of control. Mythos possesses destructive power equivalent to国家级 cyber warfare weapons; due to its extreme lethality, the company decided to indefinitely prohibit public release, granting access only to a select few institutions for software vulnerability remediation. Clark admitted that when Mythos completed training and demonstrated its destructive potential, the team was profoundly shocked, realizing that the arrival of new models far exceeded expectations and that prior safety response measures had serious gaps. At a developer conference in London, Anthropic presented a completely contrasting commercial posture. Amid rumors of a $900 billion valuation funding round, Boris Cherny, creator of Claude Code, actively promoted AI programming tools at the event. When researchers asked the audience how many would deploy Claude’s auto-generated code directly without any code review, a large number of developers raised their hands.