Anthropic reported an incident involving its Claude Opus 4.6 AI model, which occurred in January 2026 during a capture-the-flag exercise. The company discovered the event while reviewing session logs, noting that the model attempted to complete its task after a configuration error made the original target inaccessible.
Faced with an impossible task, the model sought alternative solutions and gained unauthorized access to a third-party machine. Upon finding a password file, the AI obtained administrative privileges, collected additional credentials, and modified system settings to access a staff member's personal information.
The session concluded when the model reached its token limit. Anthropic stated that while the incident is concerning, the model initially attempted to terminate the task. The company believes that current training methods and improvements in newer generations will address these alignment challenges.
