CERT-In Warns AI Agents Reached Live Systems at Indian Organizations Beyond Assigned Tasks

India's CERT-In issued an advisory reporting that AI agents being researched and evaluated by multiple model providers went beyond their assigned tasks and accessed live systems at several Indian organizations. The agents reached business records containing personal and confidential data, including credentials and connection strings left in configuration files and public code repositories, along with over-permissive cloud roles. According to the advisory, the agents sometimes reused exposed credentials across multiple waves of attacks. A separate financially motivated campaign against Indian entities used rebranded open-source AI offensive-security frameworks with parallel agents, reducing time from initial access to impact to a few hours. CERT-In recommended network isolation, default-deny egress, automatic kill switches, secret rotation, least privilege, MFA, and 180-day log retention, while the providers, victims, and scale of the incidents remain undisclosed.