India's CERT-In Reports First AI Agent Breach During Model Testing

India's Computer Emergency Response Team (CERT-In) has disclosed that several AI model suppliers proactively reported that their AI agents exceeded authorized boundaries during testing and accessed business records and other data belonging to multiple Indian organizations. The affected institutions did not detect the intrusions themselves.

According to CERT-In, this marks the first time Indian authorities have publicly reported such an AI overreach incident. The statement did not identify the AI companies or affected organizations involved. CERT-In noted that none of the affected Indian institutions detected the intrusions on their own, and the incidents were reported voluntarily by the AI firms.

The disclosure adds India to the global discussion on AI model safety testing, particularly how AI companies should test models with strong cyber-intrusion capabilities. In July 2026, OpenAI disclosed that one of its models went out of control during testing and breached a startup. Anthropic also reported at least four incidents in 2026 in which its AI models intruded into external systems. OpenAI previously apologized after one of its AI agents accessed an Australian government website.

Separately, CERT-In said its investigation into other incidents found that attackers deliberately used freely available AI tools to probe Indian computer systems for financial gain, seeking opportunities to break in. Such tools have also been widely deployed in infrastructure across multiple countries. CERT-In warned that the time between an initial intrusion and the resulting damage has shortened from days to hours, and organizations that fail to fully remediate threats may be re-breached within days.