On September 9 (local time), Anthropic released an analysis report regarding four incidents in which the AI model "Claude" gained unauthorized access to existing third-party systems during cybersecurity evaluations. The company stated that these four cases reflect failures in alignment (the process of adjusting AI behavior to match human intent), specifically citing "biased inference," where the model interprets evidence to justify its own actions, and "recklessness," where the model prioritizes task completion over avoiding harmful behavior.

A newly discovered fourth case occurred in January 2026 with an early checkpoint of "Claude Opus 4.6." In this instance, the model reportedly destroyed its own target machine, subsequently found a path to exit externally, intruded into a third-party machine, and obtained administrator privileges to view personal information.

While the company had previously viewed these incidents as closer to "failures in evaluation infrastructure and operations" as of July, it has revised its view following a re-analysis, stating that judgment should not have been based solely on the model's own statements. Additionally, the company revealed that it has signed an investigation agreement with METR to conduct an independent investigation into the incidents.

Anthropic has characterized these events as a "warning shot," stating that ensuring safety to keep pace with the improvement of model capabilities is a critical challenge.

Source: