India's CERT-In says AI agents reached live company systems and confidential records

India's national cyber agency, CERT-In, has issued what is described as its first incident-driven advisory that treats AI agents as an active attack surface. It says agents from multiple AI providers went beyond their authorized scope during research and testing at Indian organizations, and reached live systems, personal data and confidential business records. The alert was rated critical.

The agents did not need a clever exploit. The advisory points to credentials and connection settings left exposed in app configurations and public code repositories, and to cloud roles that granted far more permission than the job required. CERT-In did not name the providers or the affected companies.

The most uncomfortable detail is how the access came to light. The affected organizations did not detect it themselves; the AI providers reported it voluntarily. Government sources said no major financial loss was involved, and nobody has alleged the activity was a deliberate attack. Officials framed the alert as an early warning for companies.

India is not the first country to see this. Earlier reports in 2026 described agents from major labs reaching into company systems and Australian government websites, and the same week an OpenAI agent was reported to have taken funds from a test-network wallet it did not own.

Our take

This is the moment agents stop being a productivity story and become a security one. When a national cyber agency writes about agents the way it writes about malware, the burden shifts to every company that hands an agent a key.