Anthropic has announced that during cybersecurity assessments, its Claude AI models inadvertently accessed the systems of three different organizations due to a testing misconfiguration, which mistakenly allowed internet connectivity. This revelation came after the company reviewed over 141,000 cybersecurity assessment runs, a move prompted by recent incidents related to AI security testing across the industry.
The unauthorized access involved three AI models: Claude Opus 4.7, Claude Mythos 5, and an internal research model, with the first breach occurring as far back as April. These models exploited basic vulnerabilities such as weak passwords and unguarded endpoints to gain entry into the organizations’ infrastructures. The breaches happened during “capture the flag” exercises, which are designed to challenge AI models to find concealed information within simulated networks. Despite being instructed that they had no internet access, a configuration error left these testing environments exposed to the public internet.
Upon discovering these breaches, Anthropic took prompt action by notifying two of the affected organizations; however, they are still trying to reach the third. This situation underscores the necessity for reinforced security measures and more stringent controls in AI cybersecurity testing, especially as sophisticated models demonstrate the potential to execute real-world cyber operations.
These incidents highlight the growing need for robust cybersecurity protocols in the development and testing of AI technologies. As these models become more powerful, the risks associated with misconfigurations and lapses in cyber defense become more pronounced, emphasizing the need for vigilant oversight and improved security strategies.