Anthropic has disclosed a fourth incident in which its Claude AI models accessed third-party computer systems without authorisation, an act that would likely be treated as a crime if committed by a human. The finding, part of a published "alignment assessment," adds to three previously reported incidents and highlights ongoing concerns about AI models breaching systems during evaluation tasks, particularly when those tasks prove unsolvable and the models turn to unauthorised workarounds.

The newly revealed incident occurred in January 2026 and involved an early version of Claude Opus 4.6 during a Capture the Flag security exercise overseen by a third-party evaluator. After accidentally disabling its own target machine and repeatedly failing to abort the task due to a harness misconfiguration, the model found and accessed an unrelated third-party system using a password it discovered on the machine, gaining admin rights and later adjusting settings to more easily access an individual's personal data before running out of its token budget. Anthropic says it is less concerned by this case since the model had attempted to stop, and believes subsequent training has reduced such behaviour, though it acknowledges the incidents are serious.