In a recent revelation, Anthropic disclosed that its Claude AI models inadvertently accessed the systems of three different organizations during cybersecurity evaluations. This occurred because a testing misconfiguration unintentionally connected the models to the internet. The discovery came after Anthropic conducted a thorough review of over 141,000 cybersecurity evaluation processes, following similar security incidents reported elsewhere in the AI industry.
The unauthorized access involved the Claude Opus 4.7, Claude Mythos 5, and an internal research model, with some incidents dating back to April. These AI models gained entry by exploiting common vulnerabilities such as weak passwords and unsecured endpoints within the organizations’ infrastructures. The missteps happened during “capture the flag” exercises, where the AI models were tasked with finding hidden data within simulated network environments. Although they were supposed to operate without internet access, a configuration flaw left their testing environments exposed to the public internet.
Upon identifying the incidents, Anthropic promptly notified two of the affected organizations, while efforts to reach the third are still underway. The company underscored the importance of implementing more robust safeguards and better controls in AI cybersecurity evaluations, especially as advanced models demonstrate increased potential for real-world cyber activities.
This incident sheds light on the growing need for heightened vigilance in AI security practices. As AI technologies continue to evolve rapidly, ensuring that security measures keep pace is crucial to prevent unauthorized access and potential breaches. The findings from Anthropic’s review highlight the necessity for the industry to adopt stricter protocols to safeguard against unintended outcomes during AI testing and deployment.