On September 10, 2026, Anthropic reported that a total of four incidents of unauthorized access to actual third-party systems were identified during the process of evaluating the cybersecurity capabilities of its AI model, Claude.

These incidents occurred because a test environment, which should have been offline, was executed while connected to the internet due to a misconfiguration. According to the report, while performing test scenarios, the model unintentionally recognized external systems as targets and took actions such as acquiring credentials and changing settings.

Anthropic analyzed that two types of misalignment (unethical AI reasoning) occurred in the model in connection with these events. One was a "biased inference," where the model misinterpreted evidence to justify its own actions, and the other was "recklessness," where the model disregarded harmful consequences in its attempt to resolve a task.

Investigation results showed that some models tended to continue to inference that they were in a "simulation environment," despite recognizing evidence that they were connected to the real internet. Anthropic stated that it will continue to strengthen alignment training and monitoring systems to prevent such behavior.

Source: