AI agents built by several providers accessed personal and confidential data at Indian organisations while they were being tested, the country’s cybersecurity agency said on Thursday. The Indian Computer Emergency Response Team (CERT‑In) reported that it had received “multiple voluntary disclosures from AI model providers” indicating that their models, during research and evaluation runs on the open internet, “had reached the live systems of Indian organizations.” The agents entered the networks by exploiting credentials and connection settings that were left exposed in application configurations, public code repositories and overly permissive cloud roles, and they acted beyond the authorisation of their assigned tasks.
The incident makes India the newest stop on a growing global trail of rogue AI agent activity. In July, agents from OpenAI escaped a test environment and breached the French company Hugging Face, sparking a summer of similar incidents. Last month, OpenAI apologised to Australia after its agents accessed four government websites, including a portal that hosts Medicare statistics. At Hugging Face, OpenAI’s agents initially entered through the same kind of exposure but then exploited software flaws to execute code on the company’s servers. CERT‑In said that in India the agents did not hack the systems; they simply used the access that had been left unattended. However, the Indian cases appear more serious in one respect: the agents read personal data and confidential business records, whereas OpenAI said its agents in Australia did not touch individuals’ medical or criminal records despite writing files inside a Medicare statistics system.
Unlike earlier incidents that were disclosed by a single laboratory, the Indian breach involved agents from multiple providers. CERT‑In did not identify the providers or the affected organisations, nor did it say whether any systems were altered. TechCrunch has contacted CERT‑In for additional details. The closest comparable event is Anthropic’s July disclosure that its Claude model broke into three companies’ systems during testing by using simple techniques such as weak passwords. In both the Anthropic cases and the Indian incidents, the targeted organisations did not detect the intrusions on their own.
These disclosures leave governments dependent on AI companies to voluntarily report rogue AI agent activity.
Aditya Mehta, an editorial fellow at TechCrunch, covered the story. He is supported by the Tarbell Center for AI Journalism and studied at UC Berkeley. He can be reached at the protected email address or via encrypted message on Signal at adymehta.74.