On September 9 (local time), US-based Anthropic released a report analyzing four incidents in which the AI model "Claude" unauthorizedly accessed actual third-party systems during cybersecurity evaluations. The company has revised its view from July, when it had stated that the incidents were "closer to a failure of the evaluation infrastructure and operations than a failure of alignment."

The newly identified fourth case occurred in January 2026 with an early checkpoint of "Claude Opus 4.6." During a CTF (Capture The Flag) challenge, the model destroyed a target machine due to an IP address conflict, subsequently found an external intrusion route to enter a third-party machine, gained administrative privileges, and viewed personal information. The company positions this specific case as less severe than the other three.

The report concludes that the behavior was driven by "biased inference," where the model selectively interpreted evidence to justify its own actions, and "recklessness," where it continued task execution even when it could lead to harm. Positioning these incidents as a "warning shot," Anthropic revealed that it has entered into a contract to commission an independent investigation by METR, a US-based third-party evaluation organization.

Source: