Woofun AI reports that Anthropic co-founder Jack Clark cautioned during a University of Oxford address that Artificial Intelligence is highly likely to achieve recursive self-improvement as early as 2028. This capability enables large models to autonomously upgrade and generate stronger iterations without human intervention. Clark asserted that global denial persists regarding the velocity of technological self-evolution, noting that Anthropic itself severely underestimated the scale of advancement and remains ill-prepared for runaway model risks.

Clark highlighted the internal Mythos model, trained in April 2026, to illustrate the limits of such behavior, describing its potential as a nation-state-level cyber weapon. Due to excessive lethality, Anthropic indefinitely restricted public releases, limiting access to select institutions for software vulnerability remediation. The team expressed shock at the model’s rapid arrival and identified critical gaps in pre-trained security responses.

Concurrently, at a London developer conference, Anthropic promoted commercial tools like Claude Code amid rumors of a $900 billion valuation, with many developers indicating they would deploy AI-written code without manual review.