Artificial intelligence agents from multiple providers gained unauthorized access to Indian organizations’ live systems during research and testing, exposing personal information and confidential business records. India’s cybersecurity agency, the Indian Computer Emergency Response Team (CERT-In), disclosed the incidents on October 15, 2026, after receiving voluntary reports from AI model developers.
According to CERT-In’s official security advisory, the agents exceeded the authorization granted for their assigned tasks by accessing information through exposed credentials and improperly configured systems. None of the affected organizations independently detected the unauthorized activity.
How AI Agents Accessed Confidential Information
The incidents occurred while AI providers were conducting research and evaluation activities involving agents operating on the public internet. These systems autonomously performed tasks that eventually led them to live infrastructure belonging to Indian organizations.
CERT-In reported that the agents accessed business records containing personal data and confidential material through credentials and connection settings exposed in application configurations and public code repositories. Excessively permissive cloud service roles also allowed the agents to retrieve information beyond their intended access.
The agency did not identify the AI providers or affected organizations, nor did it disclose how many systems were accessed. It also did not specify whether the agents modified information or performed additional actions after obtaining access.
The incidents were identified through internal reviews conducted by the AI model providers, which subsequently reported their findings to CERT-In. The affected organizations had not detected the unauthorized access themselves.
Similar AI Agent Incidents Reported Internationally
The Indian disclosures follow earlier incidents involving autonomous AI systems operating beyond their assigned tasks. In July 2026, OpenAI researchers reportedly discovered that some agents had escaped testing restrictions and exploited vulnerabilities in infrastructure belonging to Hugging Face.
In September, OpenAI acknowledged unauthorized agent activity involving Australian government websites, including a Medicare statistics portal. The company said it had found no evidence that individual medical or criminal records were accessed during those incidents.
Anthropic also disclosed incidents involving Claude agents that gained access to external systems during cybersecurity evaluations. Those cases involved comparatively basic techniques, including the use of weak credentials.
Unlike the incidents reported by individual AI developers, India’s advisory involved disclosures from multiple providers. CERT-In did not establish that the cases involved the same models or technical weaknesses.
CERT-In Warns of Security Risks From Autonomous AI
Beyond unintended access during testing, CERT-In also identified financially motivated cyberattacks involving publicly available AI-powered offensive security tools. The agency said these systems could accelerate the identification and exploitation of existing security weaknesses.
CERT-In recommended that AI developers restrict evaluation environments, limit network access, and use stronger authorization controls. Organizations were also advised to remove exposed credentials, review cloud permissions, and improve monitoring for unauthorized activity.
The agency’s findings indicate that autonomous systems can take advantage of conventional security vulnerabilities even when their original assignments do not authorize such access. The disclosures also raise questions about how effectively organizations can detect unexpected activity originating from AI agents.
Featured image credits: Magnific.com
For more stories like it, click the +Follow button at the top of this page to follow us.