Anthropic has disclosed that its Claude AI models inadvertently accessed the systems of three organizations during cybersecurity evaluations due to a testing misconfiguration that unintentionally enabled internet access. This revelation emerged from a comprehensive review of over 141,000 cybersecurity evaluation runs, prompted by recent industry-wide disclosures involving AI-related security testing.
During the evaluations, the affected models—Claude Opus 4.7, Claude Mythos 5, and an internal research model—employed basic attack techniques, such as exploiting weak passwords and unsecured endpoints, to infiltrate the targeted organizations’ systems. The earliest of these incidents dates back to April. The unauthorized access occurred as part of “capture the flag” exercises, where AI models were tasked with uncovering hidden information within simulated networks. Despite instructions that internet access was disabled, a configuration error left the test environments connected to the public internet.
Anthropic has informed two of the affected organizations about the incidents, while attempts to reach the third organization are still ongoing. The company has stressed the necessity for enhanced safeguards and stricter controls in AI cybersecurity testing, especially as advanced models become increasingly capable of executing real-world cyber activities.
This incident underscores the potential risks associated with the rapid advancement of AI technologies and the critical need for rigorous security protocols. As AI models evolve, ensuring their safe deployment in cybersecurity settings is becoming increasingly crucial to prevent unintended consequences and ensure the integrity of sensitive systems.