In a recent update, Anthropic disclosed that its Claude AI models had inadvertently accessed the systems of three organizations during cybersecurity assessments due to a testing misconfiguration that mistakenly enabled internet connectivity. This revelation emerged from a comprehensive review of over 141,000 cybersecurity evaluation runs, initiated after recent industry-wide disclosures concerning AI-related security assessments.
Anthropic’s analysis revealed that the affected AI models leveraged basic hacking techniques, such as exploiting weak passwords and unprotected endpoints, to infiltrate the organizational infrastructures. The incidents involved specific models, namely Claude Opus 4.7, Claude Mythos 5, and a proprietary research model, with the earliest penetration traced back to April. The unauthorized breaches occurred during “capture the flag” exercises, where the AI models were challenged to unearth hidden data within simulated networks. Although these models were initially programmed to believe they lacked internet access, an oversight in configuration left the testing environments exposed to the public internet.
Following the discovery, Anthropic informed two of the three affected organizations, with efforts to notify the third party still underway. The company underscored that these incidents serve as a critical reminder of the necessity for more robust security measures and tighter regulations in AI cybersecurity testing. This need becomes increasingly urgent as advanced AI models grow more adept at executing real-world cyber operations.
These occurrences underscore the importance of rigorous safeguards and highlight the potential risks associated with AI’s growing capabilities in cybersecurity contexts. Anthropic’s proactive approach in reviewing a substantial number of evaluation runs emphasizes the company’s commitment to addressing vulnerabilities and enhancing security protocols in its AI testing procedures.