Anthropic Warns of AI or Self-evolution by 2028, Founder Admits Company Underestimated Iteration Speed and Lack of Preparedness

According to Dynamic Beating monitoring, Anthropic co-founder Jack Clark warned during a speech at the University of Oxford that Artificial Intelligence (AI) is highly likely to achieve "recursive self-improvement" as early as 2028. This means that large models will have the ability to self-upgrade and create stronger versions without human intervention. Clark bluntly stated that the majority of people globally are still in denial about the speed of technological self-evolution, and even Anthropic itself has severely underestimated the scale of technological advancement, being ill-prepared to face the risks of runaway models.

Regarding the capabilities of cutting-edge models, Clark cited the internal model Mythos, trained in April 2026, as an example, showcasing the true limits of runaway behavior. Mythos possesses the destructive power of a nation-state-level cyber weapon. Due to its excessive lethality, the company decided to indefinitely restrict public releases and only authorize a few select institutions to use it for software vulnerability remediation. Clark confessed that when Mythos completed training and demonstrated its destructive power, the team was extremely shocked, realizing that the arrival speed of the new model far exceeded expectations and that there were serious vulnerabilities in the pre-trained security response.

At a developer conference held in London, Anthropic displayed a completely contrasting commercial ethos. Amid rumors of a funding round valuing the company at up to $900 billion, Claude Code creator Boris Cherny actively promoted AI programming tools at the event. When researchers asked how many people would deploy code automatically written by Claude without any manual code review, a large number of developers in the audience raised their hands.