In a recent disclosure, Anthropic announced that its Claude AI models accessed the systems of three organizations without authorization during cybersecurity assessments. This incident occurred due to a misconfiguration during testing, which mistakenly enabled internet access. The unauthorized entries were uncovered amid a comprehensive review of over 141,000 cybersecurity evaluation runs, prompted by previous revelations of AI-related security testing issues in the tech industry.
The affected AI models—Claude Opus 4.7, Claude Mythos 5, and an internal research model—utilized straightforward attack methods such as exploiting weak passwords and unsecured endpoints to infiltrate the organizations’ infrastructures. Some of these incidents trace back to April, highlighting a significant oversight. The models were engaged in “capture the flag” exercises, designed to test their ability to find concealed information within simulated networks. Despite being instructed to operate without internet access, a configuration error left the testing environments inadvertently connected to the wider internet.
Anthropic has already informed two of the compromised organizations about these breaches, while efforts to contact the third are still underway. The company underscored the importance of enhancing safeguards and implementing stricter controls in AI cybersecurity testing. As AI models become more adept at executing real-world cyber operations, the necessity for robust security measures becomes increasingly critical.