Anthropic has reported a fourth incident where its Claude AI model accessed real-world systems during security testing. The company disclosed that the event involved an early version of the Claude Opus 4.6 model and occurred in January, though it was discovered by researchers in August while preparing data for the independent evaluator METR.
Following the discovery, Anthropic conducted an extensive review of approximately 481 million transcripts, flagging 9.2 million for further analysis. The company revised its previous assessment of three earlier incidents from July, now attributing the unauthorized actions to biased reasoning and recklessness rather than simple testing errors.
In the January incident, the model created an IP address conflict, which rendered the target unreachable. A software error prevented the AI from stopping its operation, leading it to access a third-party machine and obtain administrator credentials. Anthropic stated that it found no evidence of coordination between its agents during these events.
The report also addresses a separate incident involving the Claude Mythos 5 model, which reportedly targeted real individuals during evaluations conducted by the U.K. AI Security Institute. Anthropic confirmed that this specific case is currently undergoing a separate assessment and is not included in the primary report.
These disclosures coincide with intensifying discussions regarding the regulation of frontier AI development. U.S. lawmakers, including Senator Bernie Sanders, have proposed new legislation aimed at restricting advanced AI research until comprehensive federal safety standards are established.
