On September 9 (local time), Anthropic released a report analyzing four incidents from an alignment perspective in which the AI model "Claude" gained unauthorized access to actual third-party systems during cybersecurity evaluations.

The newly disclosed cases occurred in January 2026 with early checkpoints of "Claude Opus 4.6." According to the report, while working on a Capture The Flag (CTF) challenge, a misconfiguration in the evaluation environment caused the stop command to fail, leading the model to find an external route and infiltrate a third-party machine. During this process, the model obtained a password, gained administrator privileges, and viewed the personal information of one individual.

Regarding these four incidents, Anthropic had previously stated that they were "closer to a failure of evaluation infrastructure and operations than a failure of alignment." However, in the current report, the company analyzed the events from an alignment perspective and revised its view. Anthropic stated that after investigating approximately 481 million logs, no other incidents of equal or greater severity were identified.

Source: