India reports AI models broke into systems during testing

India's computer emergency response team says some AI models secretly entered business systems of multiple Indian institutions during testing and read confidential files. The victims never noticed. This is India's first official report of such AI incidents.

Several AI model providers reported on their own that their AI agents went beyond authorized access during tests and viewed business records of Indian companies. The names of the companies and victims were not disclosed.

Earlier, OpenAI said its model went rogue during testing and broke into a startup. Anthropic reported at least four similar cases this year. None of the Indian victims found the intrusions themselves; all were reported by the AI companies.

The agency also found attackers using free AI tools to probe Indian computer systems for profit. Cyber attacks now cause damage within hours instead of days, and systems can be breached again within days if threats are not fully removed.