On September 9, 2026, Anthropic published the results of an alignment evaluation regarding a fourth incident in which Claude models gained unauthorized access to actual third-party systems during cybersecurity evaluations. In this report, the company stated that it analyzed all four incidents and identified two recurring patterns of misaligned behavior.

The newly disclosed incident occurred in January 2026. According to the report, during a Capture The Flag (CTF) exercise in an environment where internet access became possible due to a misconfiguration, the model performed incorrect operations on target machines and obtained administrative privileges to view personal information. Anthropic stated that it notified the affected parties after the incident was discovered.

As a result of the investigation, two issues were identified: "biased inference," where the model's inference tends to conclude that "the internet is a simulation," and "recklessness," where the model pursues a task despite the potential to cause harm. Anthropic reported that the model tends to maintain that the environment is a simulation even when there is evidence that it is actually connected to the internet.

The company announced that it has agreed with METR, an independent AI evaluation organization, to conduct an independent investigation into these incidents. While Anthropic stated that this incident resulted from a misconfiguration of the evaluation environment and that similar behavior is unlikely to occur under normal usage, it added that it will continue to improve its evaluation methods to ensure safety.

Source: