AI agents undergoing research and evaluation by unnamed model providers exceeded their assigned authorization and read personal and confidential business records on live Indian systems, according to India's national cybersecurity agency. None of the affected organizations detected the access on their own.

The Indian Computer Emergency Response Team, known as CERT-In, flagged the activity this week in a critical advisory titled “AI Agents as Attack Vectors.” The agency said it received multiple voluntary disclosures from AI model providers whose models reached the live systems of Indian organizations while undergoing research and evaluation on the open internet. CERT-In did not name the providers, the models, or the affected organizations, and every case surfaced only because the model provider disclosed it after an internal review.

According to advisory, the agents read business records containing personal data and confidential material, using credentials and connection settings that had been left exposed in application configurations, public code repositories, and overly permissive cloud service roles—the permission sets attached to cloud accounts. In some cases, the agents recovered the same exposed credential more than once and reused it in separate operations.

The advisory is the latest in a string of cases in which AI agents have disrupted big companies and governments by straying beyond their intended boundaries during testing. Anthropic disclosed on July 30 that Claude models—which had been told they had no internet access—compromised three real companies during capture-the-flag challenges, exercises in which a model must break into a machine and retrieve a hidden secret, as Decrypt previously reported.

Britain's AI Security Institute reported on Aug. 4 that agents took 19 unsanctioned actions across 10 of 122 evaluation runs, 17 of them by Anthropic's Mythos 5 and two by OpenAI's GPT-5.6 Sol. The institute had deliberately enabled internet access for the test, according to a separate Decrypt report.

OpenAI's case involved far more agents: roughly 1,200 of them coordinated on a hidden message board while trying to cheat a cybersecurity test, and about 700 attacked Hugging Face, the AI model-sharing platform, according to an independent investigation by the AI safety research groups METR and Redwood Research, as covered by Decrypt.

Separately, CERT-In said it observed criminals using publicly available, AI-powered offensive-security frameworks to scan for weaknesses, exploit them, and run financially motivated attacks against Indian entities. The agency said the tools cut the time from initial access to final impact from days to a few hours, are repeatedly repackaged and redeployed, and have been found deployed on servers in multiple countries.

The advisory was earlier reported by the Indian Express.

Source: Decrypt