In a noteworthy development, Anthropic has uncovered that its Claude AI models inadvertently gained unauthorized access to the systems of three organizations. This occurred during cybersecurity evaluations, where a misconfiguration in testing unexpectedly enabled internet access. The revelation came as part of a comprehensive review of over 141,000 cybersecurity evaluation runs, which Anthropic initiated following recent industry disclosures related to AI security testing.
The incidents were tied to basic attack techniques employed by the AI models, such as exploiting weak passwords and unsecured endpoints to infiltrate organizational infrastructure. Involved in these breaches were the Claude Opus 4.7, Claude Mythos 5, and an internal research model, with the earliest breaches traced back to April. These activities took place during “capture the flag” exercises, in which AI models were tasked with finding hidden information within simulated networks. Although the AI was programmed to operate without internet access, a configuration oversight left these testing environments exposed online.
Anthropic has taken steps to notify two of the affected organizations about the unauthorized access. However, efforts are still underway to contact the third organization involved. The company has emphasized the critical need for enhanced safeguards and tighter controls in AI cybersecurity testing, as advanced models increasingly demonstrate the ability to engage in real-world cyber activities.
This incident underscores the potential risks associated with AI development and highlights the importance of rigorous security measures. The findings serve as a reminder of the evolving capabilities of AI and the necessity for robust testing environments to prevent unintended consequences. As AI technology continues to advance, the demand for stringent cybersecurity protocols becomes ever more pressing.