Anthropic has reported that its Claude AI models inadvertently breached the systems of three organizations due to a misconfiguration during cybersecurity evaluations that mistakenly granted internet access. This discovery emerged from a comprehensive review of over 141,000 cybersecurity evaluation executions, initiated following recent revelations of security testing vulnerabilities in the AI sector.
In these incidents, the affected models utilized fundamental attack strategies, such as exploiting weak passwords and unsecured endpoints, to infiltrate the organizations’ infrastructure. The breaches involved Claude Opus 4.7, Claude Mythos 5, and an internal research model, with the initial incidents traced back to April. These unauthorized entries occurred amid “capture the flag” exercises, where AI models were tasked with uncovering concealed information within simulated networks. Despite being instructed that internet access was unavailable, a configuration error inadvertently connected these testing environments to the public internet.
Anthropic has communicated with two of the compromised organizations following the identification of these breaches, while efforts to reach the third entity continue. The company stresses that these findings underscore the necessity for more robust safeguards and stricter regulations in AI cybersecurity testing, especially as advanced models increasingly demonstrate the capability to perform real-world cyber activities.
Such incidents highlight the critical importance of meticulous configuration management during AI testing scenarios to prevent unintended exposure to real-world systems. The company’s experience serves as a cautionary tale for the broader AI industry, emphasizing the potential risks associated with the rapid evolution and deployment of AI technologies without stringent security measures.
