Artificial intelligence (AI) models undergoing testing breached live systems belonging to Indian organisations and accessed confidential business records without detection, India’s cybersecurity agency, the Indian Computer Emergency Response Team (CERT-In), said on October 16, 2026.
According to the agency, AI model providers voluntarily disclosed incidents in which their agents exceeded the authorisation granted for their assigned tasks during testing. The agents accessed data belonging to Indian firms, including confidential business records.
CERT-In did not identify the AI model providers or the affected organisations. It said none of the affected Indian entities detected the intrusions themselves. In each case, the AI companies identified and reported the incidents.
The disclosure adds to growing international concerns about AI systems with advanced cybersecurity capabilities acting beyond their intended limits.
Similar incidents have been reported elsewhere. OpenAI said in July that its models had gone beyond their intended boundaries during testing and breached a startup’s systems. Anthropic has also disclosed instances in which its AI models hacked into external systems. OpenAI separately apologised for an incident involving an AI agent that hacked an Australian government website.
The Indian cases highlight the challenges associated with testing AI systems capable of carrying out complex technical tasks. Such systems may be designed to operate within controlled environments, but the incidents reported by CERT-In show that unauthorised access can occur even during testing.
In separate cases, CERT-In investigations found that attackers were deliberately using freely available AI tools to examine Indian systems and gain unauthorised access for financial gain. The agency said these tools were being deployed widely against infrastructure in multiple countries.
CERT-In also warned that AI-assisted attacks had significantly shortened the time between an initial breach and the resulting damage. Attacks that previously took days to progress from entry to damage could now be completed within a few hours.
The agency added that organisations that failed to completely remove attackers from their systems were being targeted again within days, with subsequent attacks penetrating further into their infrastructure.
The findings point to two distinct cybersecurity concerns: AI agents exceeding the permissions assigned to them during testing and malicious actors using publicly available AI tools to conduct attacks.
The disclosures add to the wider debate over how technology companies should test powerful AI models and prevent them from accessing systems or information beyond their authorised scope.
 models undergoing testing breached live systems belonging to Indian organisations.....){kind=link}