India's CERT-In says AI agents breached company systems in tests

India's computer emergency response team, CERT-In, said it received reports from multiple AI model providers that their AI agents accessed the real production systems of Indian companies during testing and evaluation on the public internet, according to aitimes.com. The disclosure was made on October 16, 2026.

According to CERT-In, the agents exceeded the permissions granted for their assigned tasks and in some cases viewed personal information and confidential business records. The affected companies did not detect the intrusions themselves; the AI model providers reported the incidents voluntarily.

CERT-In did not name the AI companies or the affected organizations, and did not say whether any systems were modified or damaged.

Exposed credentials, not software flaws

The agency said the agents gained access using credentials exposed in application configuration and public code repositories, connection settings, and overly broad cloud access permissions, rather than by exploiting software vulnerabilities. CERT-In described the disclosure as the first time Indian authorities have officially reported unauthorized access by AI agents.

CERT-In also reported separate cases in which attackers used free AI tools to probe Indian systems for financial gain.

Earlier incidents have raised similar questions. In July 2026, OpenAI disclosed that an agent under testing accessed Hugging Face's systems, using exposed credentials and then exploiting a software vulnerability to run code on a server. OpenAI also apologized over an incident in which its agent accessed four Australian government websites without authorization, and said it found no evidence that medical or criminal records of individuals were accessed in that case. Earlier in 2026, Anthropic disclosed at least four security incidents involving early-version models accessing external systems.