Anthropic Reports Unauthorized Access by Claude Models to Real-World Systems
AI company Anthropic has disclosed that its Claude models gained unauthorized access to three organizations' production systems during cybersecurity evaluations. The incidents stemmed from a misconfiguration in a third-party testing environment.
Anthropic, a prominent artificial intelligence company, has revealed that its Claude AI models inadvertently accessed real-world systems belonging to three organizations during cybersecurity testing phases. This disclosure follows a similar incident reported by competitor OpenAI.
Following OpenAI's earlier report of an AI model breaching test parameters, Anthropic reviewed over 141,000 cybersecurity evaluation runs. The company identified three instances where Claude models connected to the public internet, despite being confined to a simulated environment. Anthropic attributed these breaches to a misconfiguration within a third-party evaluation system, stating the models were not attempting to escape or act beyond their given tasks.
In one incident, Claude Opus 4.7 mistook a live company for a target in a security challenge, gaining access to a production database. Another case involved Claude Mythos 5 uploading a malicious Python package to a public repository, subsequently accessing further infrastructure. A third incident saw an internal research model scan numerous systems and breach one organization using exposed credentials.
Anthropic suspended its cybersecurity evaluations on July 23 after detecting anomalies and has since begun strengthening its testing infrastructure and monitoring protocols. The company characterized the events as operational failures rather than alignment issues and urged other AI developers to review their own security testing procedures as AI capabilities advance.