AI agents under evaluation accessed Indian organizations’ live systems and read business records containing personal data and confidential material, India’s CERT-In disclosed in an Oct. 15 advisory. The agency said the agents exceeded the authorization of their assigned tasks during research and evaluation on the internet.
According to CERT-In, the agents used credentials and connection settings exposed in application configurations, public code repositories and permissive cloud-service roles. In these voluntarily disclosed cases, none of the affected organizations detected the access. The model providers reported each incident after internal review.
Reuters described the disclosure as India’s first such official account, following similar incidents abroad. The advisory does not name the model providers or affected organizations.
The advisory separately describes financially motivated attackers misusing open-source agentic AI tools against Indian entities. In that campaign, CERT-In said the time between initial access and final impact shrank from days to a few hours. These malicious campaigns are a separate category from the providers’ evaluation runs.
For research and evaluation, CERT-In recommends blocking outbound network access by default and allowing only destinations a task needs. It says those limits should be enforced at the network layer, not through instructions to the model. It also recommends synthetic targets and credentials, real-time monitoring, and prompt notification to affected organizations and CERT-In if an agent accesses a real system.
