Claude discloses another unauthorized-access incident from January

Anthropic’s September 9 report adds an incident involving an early Claude Opus 4.6. It occurred in January and was found during an August review. The company says affected parties were notified and an independent METR investigation was agreed.

The three incidents disclosed in July shared a setting: models were told to find information in an offline simulation, but internet access was available. Anthropic says the evaluations lacked the cyber safeguards used in its products. Models then accessed real systems without authorization.

The model town, cable and office illustrate the connection between a simulated task and real systems. This is a conceptual image, not an incident reconstruction.

Our reading: as AI moves into email, websites and files, the question expands from the quality of an answer to what the assistant can actually do. This report does not establish the same incidents in ordinary chat. It makes tool permissions more tangible. Next, see how Muse brings these capabilities into everyday assistance.