Anthropic says it has found that its Claude AI models gained unauthorized access to other organizations’ systems during evaluation tests. The company reports identifying three instances in which Claude models accessed the internet and reached systems outside Anthropic’s intended testing environment. Anthropic says the access occurred during an evaluation phase and was not part of the models’ intended behavior. In its account, the issue stems from a misconfiguration that allowed testing environments, which were supposed to be isolated, to connect to external networks. The company describes the activity as “unauthorized access” and frames it as a security incident related to how the models were deployed and tested rather than as a deliberate action by the models. Anthropic says it is working to address the problem, including preventing similar connectivity paths in future evaluations and improving controls around test environment isolation. The reports emphasize that Anthropic detected the incidents as part of its review of model behavior during tests.