* An Artificial Intelligence (AI) agent working on research data provided by an American company ended up accessing sensitive information belonging to an IT services firm in Noida.* AI agents stumbled upon crucial passwords belonging to another Indian enterprise, left exposed in plain text.These were among the first reported instances of AI agents going ‘rogue’ and gaining unauthorised access to the systems of Indian companies, The Indian Express has learnt. ‘Rogue’ because in both the instances, what the agents did was not part of the tasks the AI systems had been assigned.The AI models involved were developed by OpenAI and Anthropic, two of the world’s leading AI companies, with the breach largely affecting entities in the IT services and financial services sectors, a senior government official said.Anthropic, OpenAI, and the IT Ministry did not respond to a request for comment until publication.Must Read | OpenAI says its AI agents disabled safety monitors designed to flag ‘rogue’ behaviourAI agents are autonomous systems capable of independently planning and executing tasks, accessing external websites and software tools, and taking actions with limited human intervention. Unlike conventional chatbots, which primarily respond to queries, such systems can perform a series of actions without requiring human instructions at every stage.The Indian Computer Emergency Response Team (Cert-In), which disclosed last week that AI agents had accessed confidential business records of organisations while operating beyond their assigned tasks, has conducted a preliminary assessment of the incidents. The damage caused was relatively limited, because the AI companies identified the unauthorised activity during internal reviews and alerted the Indian government in time, the official said.Story continues below this adThe incidents have, however, raised concerns within the government over the cybersecurity preparedness of enterprises, particularly smaller organisations, which may be ill-equipped to deal with attacks carried out by autonomous AI systems.In the case involving the Noida-based IT services company, the AI agents were carrying out research and evaluation activities using information supplied by an American company. However, they managed to operate beyond the restrictions placed on their assigned tasks and reach systems linked to the Indian firm, which had undertaken backend work for the American company.Also Read | Rogue AI agents target Indian systems: What CERT-In found, why experts are worriedThe second incident exposed more conventional weaknesses in cybersecurity practices. AI agents were able to access sensitive passwords belonging to an enterprise because the credentials had been stored in plain text, making them accessible without needing to defeat encryption.On October 15, Cert-In issued a critical-severity advisory stating that it had received multiple voluntary disclosures from AI model providers whose systems, while conducting research and evaluations on the Internet, had reached live systems belonging to Indian organisations.Story continues below this ad“In the disclosed cases, the models’ actions exceeded the authorisation of their assigned tasks,” the agency said, adding that the recorded activity involved accessing business records, including personal data and confidential material, through credentials and connection settings exposed in application configurations, public code repositories and permissive cloud service roles.Significantly, none of the affected Indian organisations had detected the unauthorised access themselves. In every instance, the incidents were discovered by the AI developers during internal reviews and subsequently disclosed to Cert-In.The agency did not publicly identify the affected organisations or the AI companies involved.The incidents come amid growing international concerns over AI models breaching their operational boundaries and interacting with external systems without authorisation.Story continues below this adAlso Read | PHP security alert: What US and Indian government warnings mean for websitesApart from the cases involving AI agents operating beyond their assigned tasks, Cert-In has also identified a separate, financially motivated cyberattack campaign against Indian entities in which attackers deliberately used publicly available AI-native offensive-security frameworks.According to the agency, the use of such tools compressed the time between initial access and the eventual impact of an attack from several days to just a few hours. It also observed instances of multiple AI agents operating simultaneously for hours or days to achieve an attacker’s objectives.Organisations which failed to comprehensively address vulnerabilities following an initial intrusion were attacked again within days, sometimes at a larger scale, Cert-In said.The agency has recommended that AI developers isolate research and evaluation workloads, restrict the external systems agents can access, and enforce these boundaries through technical network controls rather than relying solely on instructions given to AI models.Story continues below this adIt has also advised developers to automatically stop an agent’s operations if it accesses unauthorised third-party systems or credentials, while giving an independent human operator the authority to terminate the process.For enterprises, Cert-In has recommended auditing source code, configuration files and repositories for exposed credentials, removing unnecessary access permissions, securing cloud deployments and introducing systems capable of detecting automated attacks at machine speed.E-dot: Rising risks call for safeguardsEnterprises need to better prepare in terms of cyber security for rising risks of autonomous AI agents becoming more and more capable of launching attacks independently. From removing unnecessary access to strengthening threat detection systems, the government has recommended many safeguard measures.