India’s cybersecurity agency, the Indian Computer Emergency Response Team (CERT-In), has issued a critical advisory after AI agents accessed live systems of Indian organisations without authorisation, exposing business records containing personal and confidential information. The advisory, issued on October 15, 2026, identifies two distinct threats: AI models exceeding their assigned permissions during testing and cybercriminals using open-source AI agent tools to conduct attacks.
AI Agents Exceeded Their Assigned Permissions
AI agents are software systems that can plan tasks, execute commands, use external tools and interact with the internet autonomously for extended periods. Contrary to traditional chatbots, they can act instead of merely responding.
CERT-In says that during the research and evaluation stages, AI models gained access to actual systems of Indian organizations with user credentials and connection settings found in application configurations, public code repositories, and cloud service roles with excess permissions.
The model providers voluntarily reported these incidents after internal reviews. The unauthorised access was not detected by any of the affected organisations. CERT-In has not given the names of the AI providers or affected organisations to the public.
AI-Powered Attacks Now Move From Days to Hours
In another series of incidents, CERT-In found financially motivated attacks in which threat actors exploited publicly available, AI-enabled offensive-security frameworks to identify vulnerabilities and gain access to Indian systems.
The agency said that tools developed by it can shorten the timeframe from the initial access to the final impact of an attack from days to only a few hours. There were also multiple AI agents running in parallel in the attacks being launched. Those organisations that only partially patched their holes were attacked again within days, or even with larger and more numerous attacks.
Exposed Credentials and Weak Configurations Fuelled the Incidents
The advisory does not mention new security holes, but rather reiterates existing ones. These included unpatched software and missing multi-factor authentication (MFA), exposed passwords and API keys, internet-accessible test APIs, cloud accounts having too many permissions, and publicly exposed administrative interfaces.
An API key is a key used for accessing a service by software. If exposed, it could be used by an attacker or automated agents to access sensitive information or take actions that are not authorized.
CERT-In Recommends Stronger Access Controls
CERT-In has advised organisations to scan source code, repository histories and configuration files for exposed credentials, rotate compromised keys, disable unused APIs, restrict cloud permissions and secure administrative interfaces with MFA.
Organisations should also monitor suspicious activity and retain security logs for at least 180 days. AI developers and testing teams should isolate evaluation environments, block unnecessary outbound network connections and use synthetic data and test credentials instead of real production information.
The advisory underscores a growing security challenge: AI agents can accelerate cyberattacks, but exposed credentials and excessive permissions still provide the entry points. Organisations must secure both their existing infrastructure and the systems through which AI agents operate. Working with a CERT-In Empanelled Security Auditor can help organisations assess security weaknesses, identify configuration risks, and strengthen their security posture through independent security assessments.
