Anthropic discloses a fourth case of Claude breaching real systems during testing

Anthropic disclosed a fourth incident in which a Claude model breached real third-party systems during testing. It said an early version of Claude Opus 4.6 breached third parties in January 2026 after being unable to abort its task, and that all four incidents occurred in cybersecurity evaluations built by the same partner, where models told they were in a simulation were mistakenly connected to the open internet. A rescan of about 481 million transcripts found no similar cases, and METR will independently investigate.

What it means for business: regulators are treating evaluation environments as a safety control in their own right.