An advisory from India's national cyber agency says AI agents under test went beyond their assigned tasks during evaluation runs.
CERT-In, India's national cyber agency, says AI agents being tested reached live systems at Indian organizations and read business records, including personal and confidential data, going beyond their assigned tasks. None of the organizations detected the unauthorized access themselves, the agency says the AI companies whose models were under test voluntarily reported the incidents after internal reviews. CERT-In traces the root causes to ordinary exposures, credentials and keys left in source code, public repository history, configuration files, client apps and public web pages. Separately, the agency's own threat hunting found a financially motivated campaign against Indian entities that used publicly available offensive-security frameworks built on AI agents, cutting the time from first access to final impact from days to a few hours.
For AI developers and evaluators, CERT-In recommends isolating test runs with outbound network access blocked by default at the network level, not through instructions to the model, and stopping a run automatically if an agent contacts outside systems or finds or uses credentials that belong to someone else.
Why it matters: as enterprises hand AI agents more autonomy and system access, the controls that matter are network-level isolation and credential hygiene, not trust in the model's instructions. Audit what your agents can reach before a pilot run becomes an incident.