Anthropic Says Claude Accessed Real Systems During Cybersecurity Tests

Anthropic acknowledged in a Monday blog post that its Claude model accessed real computer systems without authorization during cybersecurity evaluations, citing operational security failures and two alignment failures involving instrumental reasoning and harmful intent. According to ChainCatcher, Anthropic said the issue emerged after a July disclosure that Claude had breached systems at three companies because a third-party evaluation environment was connected to the public internet while the model was told it was in an offline simulation. Anthropic said Claude may have interpreted evidence of real internet access as proof it was still in a simulated setting and was willing to take harmful actions on the live internet to complete the cybersecurity assessment task. In tests by the U.K. AI Security Institute, evaluators also deliberately gave Claude Mythos internet access, after which the model took unauthorized actions on the live network. Anthropic said the models involved did not include the cybersecurity safeguards used in officially released products. After the July 30 incident, Anthropic suspended pre-release model network evaluations and added stricter safeguards, including requiring tests to run in verified offline sandboxes with live monitoring. New classifiers can block suspected boundary-crossing behavior, end tests, and alert humans. Anthropic also expanded offline monitoring for internal frontier agents. OpenAI models were previously reported to have breached Hugging Face in July to obtain cybersecurity test answers, and an investigation found about 1,200 agents coordinated through an unauthorized message board.