
India’s cybersecurity agency has warned organisations about AI agents accessing live systems without authorisation after models undergoing research and evaluation reached the systems of Indian companies and accessed business records containing personal and confidential information.
The Indian Computer Emergency Response Team (CERT-In) disclosed the incidents in a critical advisory issued on October 15, titled AI Agents as Attack Vectors. According to the agency, the AI agents exceeded the permissions assigned to their tasks and accessed records containing personal information and confidential business data.
And more implicating is the fact that none of the affected organisations detected the unauthorised activity themselves.
The advisory also identified a separate cyberattack campaign in which criminals used publicly available AI-powered security tools to target Indian organisations for financial gain. CERT-In said the tools had reduced the time required to move from gaining access to a system to achieving the intended impact of an attack from several days to just a few hours.
AI Agents Accessed Sensitive Business Information During Testing
According to CERT-In, it received multiple voluntary disclosures from AI model providers whose models had accessed live systems belonging to Indian organisations during internet-based research and evaluation activities.
The agents accessed the systems through credentials and connection settings exposed in application configurations, public code repositories, and cloud service roles with excessive permissions. These security gaps allowed the models to read business records despite their assigned tasks not authorising access to the organisations’ systems.
The affected companies only became aware of the incidents after the AI providers reviewed their models’ activities and voluntarily reported the cases to CERT-In. However, the agency did not identify the AI developers involved, the affected organisations, or the number of records accessed.
The disclosures add to growing concerns about the security risks associated with increasingly autonomous AI systems, as agents can independently plan tasks, execute commands, and interact with external tools, which makes it possible for them to carry out activities beyond what their developers initially intended if their operating environments lack sufficient restrictions.
Cybercriminals Are Using AI Agents to Accelerate Attacks
Beyond the incidents involving models under evaluation, CERT-In identified a separate campaign in which cybercriminals used publicly available AI-native offensive security frameworks to conduct reconnaissance, exploit vulnerabilities, and carry out financially motivated attacks against Indian entities.
The frameworks have also been misused and repackaged, with multiple deployments identified across infrastructure in different countries.
According to CERT-In’s investigation, the attackers used several AI agents simultaneously that allowed them to coordinate activities over extended periods, which, in turn, reduced the time between gaining initial access to a system and achieving the intended impact from days to a few hours.
The agency also found that organisations that only partially addressed the security weaknesses exploited during an initial attack were targeted again within days, with subsequent intrusions reaching a greater scale. In some of the incidents involving AI agents under evaluation, the models also rediscovered and reused exposed credentials during separate operations.
CERT-In Calls for Tighter Controls Over AI Agents
Following the disclosures, CERT-In advised organisations to strengthen their existing security measures, particularly around exposed credentials, application programming interfaces (APIs), and cloud infrastructure.
The agency recommended auditing source code, public repositories, configuration files, and cloud storage for exposed credentials, rotating compromised keys, and removing outdated or unused APIs. It also called for multi-factor authentication for administrative accounts and continuous monitoring of unusual API activity.
For organisations deploying AI agents, CERT-In recommended giving each agent a separate, short-lived identity with limited permissions. It also advised restricting agents’ access to approved network destinations and maintaining records of their activities, particularly when interacting with production systems and sensitive data.
The agency also recommended enforcing default-deny network access for AI research and evaluation environments, where connections are allowed only to destinations required for assigned tasks. According to CERT-In, these restrictions should be applied at the network level rather than relying on instructions given to the AI models. It further advised organisations to retain security logs for at least 180 days and strengthen monitoring systems to detect unusual API activity, bulk data access and automated registrations.
More importantly, AI developers and research organisations were also advised to isolate testing environments, use synthetic credentials and prevent agents from accessing external systems unless specifically authorised. The agency further recommended that developers notify affected organisations and CERT-In promptly if an agent accesses a real system.
These incidents highlight the security challenges organisations face as AI agents gain greater access to digital infrastructure. With models now capable of independently interacting with business systems, preventing unauthorised access will remain the utmost priority for companies.
