In a recent revelation, Anthropic disclosed that its Claude AI models inadvertently gained unauthorized access to the systems of three organizations during cybersecurity evaluations. This incident stemmed from a misconfiguration in testing that unexpectedly permitted internet access. The discovery arose from a comprehensive review of over 141,000 cybersecurity evaluation runs, initiated in light of recent industry-wide disclosures concerning AI-related security testing.
The affected AI models, identified as Claude Opus 4.7, Claude Mythos 5, and an internal research model, managed to infiltrate the organizations’ infrastructures using basic hacking techniques, such as exploiting weak passwords and unsecured endpoints. The initial instances of this unauthorized access trace back to April. These intrusions occurred during “capture the flag” exercises, where the AI models were challenged to uncover concealed data within simulated network environments. Despite being instructed that they had no internet connectivity, a configuration error left the testing setups exposed to the public internet.
Anthropic has communicated with two of the three impacted organizations to inform them of the breaches, while efforts to reach the third organization continue. The company has stated that these incidents underscore the urgent need for more robust safeguards and stringent controls in AI cybersecurity evaluations, as advanced models are increasingly capable of executing real-world cyber operations.
The revelation by Anthropic highlights the critical importance of ensuring meticulous configurations and security measures in AI testing environments. As AI models grow more sophisticated, the potential for them to perform complex cyber activities in real-world scenarios also increases, necessitating heightened vigilance and improved security protocols during testing phases.