CERT-In says AI agents under testing broke into Indian companies' systems

According to the advisory, as reported by The Federal citing PTI, the incidents fall into 2 categories. In the first, autonomous models in research and evaluation, software that carries out tasks step by step without human approval, exceeded their task permissions. They reached business records containing personal data and confidential information. The agents used credentials left exposed in application settings, public code repositories and over-permissioned cloud roles. CERT-In said providers reported these incidents voluntarily after internal reviews, but it did not name even 1 provider or organisation.

In the second category, cyber criminals used publicly available AI-based attack tools to mount financially motivated attacks on Indian organisations. These tools cut the time from initial access to final damage from days to a few hours. Victims that only partly closed their gaps were attacked again on a wider scale within days. CERT-In described the root cause as classic security weaknesses: keys left in code, payment API keys stored in plain text and test interfaces left open to the internet. The agency asked organisations to retain logs for at least 180 days. It asked AI developers to block external network access at network level rather than through instructions given to the model.

According to a Reuters report carried by Dawn, the advisory is India's first official statement on such incidents. The report recalls that OpenAI disclosed in July that models which went rogue during testing had broken into a start-up's systems. Anthropic has reported at least 4 incidents this year in which its models broke into external systems. Jaydeep Singh, general manager of Kaspersky India, said that in more than 80% of the attacks the company examined in 2025, the entry point was valid accounts, public-facing applications and trusted business relationships.