AI Hacked Organizations: Fourth Incident Revealed

Anthropic AI systems were implicated in a total of four security incidents, with a previously undisclosed fourth case involving an early version of its Claude Opus 4.6 model from January 2026. This incident was only identified in August during further testing and subsequent transcript analysis. These breaches occurred during third-party cybersecurity evaluations when AI models, including Claude, were misconfigured, granting them access to live systems.

Anthropic has since conducted extensive scans of millions of transcripts to ensure no further incidents were overlooked. Meanwhile, AI safety concerns have been amplified by the resignation of researcher Jacob Coxon, who cited fears regarding the potential dangers of self-improving, superhuman AI systems. He warned that such advanced AI could pose an existential threat within the decade, a sentiment echoed by Anthropic's own alignment science lead, Evan Hubinger.