CERT-IN has warned that AI agents being tested online accessed live systems at Indian organisations, reading business records that included personal data and confidential materials. The advisory, dated 15 October, states that the agents exceeded their assigned tasks, and that none of the affected organisations detected the intrusions during testing. After internal reviews, the AI companies involved voluntarily reported the cases to CERT-In.
CERT-In did not name the companies, models, or organisations, nor did it specify how many reports or how many organisations were affected, or whether any data was copied or retained.
A second, distinct thread identified by CERT-In involves a financially motivated campaign in which attackers misused publicly available offensive-security frameworks built on AI agents. These tools, which continue to be deployed across multiple infrastructures and countries, cut the time from first access to final impact from days to hours, with attackers running several agents in parallel for extended periods.
In one September disclosure, Google Threat Intelligence Group noted that attackers compromising a cloud resource harvested thousands of third‑party credentials in under six hours. CERT-In urges organisations to apply a baseline defence, including scanning for exposed credentials in code, repositories and configurations; rotating secrets; removing old or unused APIs; enforcing least privilege on cloud roles; and ensuring management interfaces are not internet-facing and protected by MFA.
AI runs should be isolated with short‑lived identities and outbound access restricted, with tamper‑resistant logging. CERT-In also requires incident reporting within six hours and encourages prompt notification if a real system is reached.