AI model providers have voluntarily disclosed incidents in which their agents “exceeded the authorisation of their assigned tasks” during test runs and accessed data belonging to Indian companies, including business records, the Indian Computer Emergency Response Team (CERT-In) said late on Thursday.
The agency did not identify the AI providers or the affected organisations.
The disclosure adds India to a growing global debate over how AI companies test models equipped with advanced hacking capabilities.
In July, OpenAI reported that its models had gone rogue during testing and breached a startup. Anthropic has also disclosed at least four incidents this year in which its AI models hacked external systems.
OpenAI has separately apologised after one of its rogue AI agents hacked an Australian government website.
According to CERT-In, none of the affected Indian organisations detected the intrusions independently. In every case, the AI companies themselves reported the incidents to the authorities.
In a separate set of cases, the government agency’s own investigations identified attackers deliberately using freely available AI tools to probe Indian systems and gain unauthorised access for financial gain. Such tools, it said, were being widely deployed against infrastructure across multiple countries.
CERT-In warned that attacks that previously took days to progress from initial access to final damage were now being completed within a few hours. It added that victims who failed to fully remove attackers from their systems were targeted again within days, with subsequent attacks penetrating deeper into their networks.
