Rogue AI agents breach Indian companies' systems, access confidential data undetected
India's cybersecurity agency says AI models exceeded authorised tasks during testing, while criminals are using AI tools to accelerate cyberattacks
BENGALURU: Artificial intelligence (AI) agents undergoing testing gained unauthorised access to Indian companies' live systems and confidential business records without being detected, India's cybersecurity agency has disclosed.
The Indian Computer Emergency Response Team (CERT-In) said AI model developers voluntarily reported incidents in which their agents exceeded the permissions granted for assigned tasks and accessed sensitive corporate information.
According to the agency's statement issued on Thursday, none of the affected organisations detected the breaches independently. The incidents were identified and reported by the AI companies conducting the tests.
CERT-In did not disclose the identities of the AI developers or the organisations whose systems were accessed.
The disclosure marks India's first official account of such incidents, following similar reports involving AI models accessing external systems without authorisation in other countries.
OpenAI disclosed in July that its AI models had breached a start-up's systems during testing, while Anthropic has reported at least four incidents involving models accessing external systems this year.
OpenAI has also apologised for an incident in which one of its AI agents accessed an Australian government website without authorisation.
Separately, CERT-In said its investigations had uncovered cases in which cybercriminals deliberately used publicly available AI tools to identify vulnerabilities and breach computer systems for financial gain.
The agency found that such tools were being deployed against infrastructure across multiple countries, significantly accelerating the pace of cyberattacks.
According to CERT-In, attacks that previously required several days from initial intrusion to causing damage could now be completed within hours.
It also warned that organisations failing to fully remove malicious access following an attack were being targeted again within days, often through more extensive intrusions.