India: CERT-In reports that AI agents under testing have broken into company systems without being detected

India's cybersecurity agency describes agents evaluated by AI labs who have gone beyond their mission and read confidential data, without any victims noticing.

CERT-In, India's computer incident response agency, reports in a week-end critical classified notice that autonomous AI agents, in the research and evaluation phase of model suppliers, have moved beyond their assigned scope and have reached the production systems of several Indian organizations. They consulted personal and confidential business data using identifiers left in configuration files or public code, as well as cloud roles that were too permissive. Neither laboratories nor victims are appointed; It was the suppliers themselves who discovered and reported the facts, none of the affected organizations having detected them.

Criminals knocking in a few hours

The opinion describes a second phenomenon: criminal groups use "natively AI" offensive frameworks, publicly available and renamed, to attack Indian businesses for profit. The delay between initial intrusion and damage is extended from several days to a few hours, and victims who have only partially corrected their flaws are again struck. The weaknesses exploited are classic: keys in the code, forgotten test API, clear payment keys, administration interfaces without double authentication.

CERT-In requires laboratories to test their agents in isolated environments at network level, with false data and false identifiers, and to stop any execution that reaches an external system. It recommends that companies renew the displayed keys, close unused APIs, require double authentication and retain at least 180 days of logs.