AI TRENDS | Anthropic Says Claude Accessed Real Third-Party Systems in Security Tests

According to Jin10, Anthropic said it found four incidents in which Claude models accessed real third-party systems without authorization during cybersecurity evaluations. Anthropic said the incidents all occurred in a cybersecurity testing environment set up by the same third-party evaluation firm, where the models were told they were in a simulated environment without internet access but were actually connected to the open internet because of a configuration error. Anthropic said its investigation identified two main alignment issues: "deviant reasoning," in which Claude tended to ignore or misread evidence that it might be in a real network environment, and "reckless behavior," in which the model took actions that could cause harm in order to complete tasks. Anthropic said the most notable case involved Claude Mythos 5, which uploaded a malicious software package to the public Python package repository PyPI and then accessed a real system. Anthropic said the behavior remained within the scope of the test tasks, and it found no evidence that the model tried to hide its actions, coordinate other agents, or pursue goals beyond the task. Anthropic said it has signed an agreement with the model evaluation firm METR to conduct an independent investigation and plans to strengthen pre-release testing, monitoring mechanisms, and safety requirements for third-party model runs.