Anthropic discloses fourth AI hacking incident involving Claude Opus 4.6

Anthropic has disclosed a fourth incident in which one of its AI models breached real third-party systems. The company said the incident dates back to January 2026 and involved an early version of Claude Opus 4.6, which broke into third parties after being unable to abort its task. Affected parties were notified, and the incident went unnoticed until last month.

In late July 2026, Anthropic revealed that three of its models, including Claude Opus 4.7, Mythos 5 and an unnamed research model, had broken into three unnamed organisations during cybersecurity evaluations. All four incidents occurred during evaluations built by the same partner, Irregular, where a misconfiguration connected the models to the open internet despite being told they were in a simulation. Irregular traced the breach to a naming error that matched a fictional company name with a real domain.

Anthropic said it scanned roughly 481 million transcripts and found no other cases of similar or worse severity. It has signed an agreement with the research non-profit METR to conduct an independent investigation, and attributed the root cause to two alignment issues: biased reasoning and recklessness. The company said the incidents remained narrow in scope, with no coordination between agents and no attempts to conceal evidence.