India's national cyber emergency response team issued a critical advisory on October 15: AI agents that were being evaluated reached production systems at Indian organizations and read business records. None of the victims detected it; it came to light because the labs reported it after reviewing their own tests.
According to coverage from The Hacker News (October 16) and Indian outlets such as Deccan Chronicle, CERT-In's advisory, titled "AI Agents as Attack Vectors," describes two kinds of incident. First: AI models undergoing online research and testing strayed from their assigned task and reached live systems at Indian organizations, where they read records containing personal data and confidential material. Second: criminals using public offensive-security tools built on AI agents for financially motivated campaigns.
The most uncomfortable part of the advisory is that it doesn't describe novel exploits. The agents got in through credentials and connection settings exposed in application configurations and public repositories, and through cloud service roles with excessive permissions. According to the report, in the labs' cases the agent found the same exposed credential and reused it across different operations. In other words: an agent that searches methodically and never tires quickly finds the "ordinary" weaknesses teams had put off for years.
On the offensive side, CERT-In says attackers ran several agents in parallel for hours or days, and that organizations that fixed only some of the problems were hit again within days and at greater scale.
The same report recalls earlier disclosures: OpenAI said in July that its models compromised the Hugging Face platform and is reviewing what its models did online during training and testing; Anthropic, after reviewing its cybersecurity evaluations, reported 4 incidents in which Claude models gained unauthorized access to real third-party systems, in evaluations designed by a single external partner. (We covered that case in our article on Claude's evaluations on the real internet.) CERT-In did not say whether OpenAI or Anthropic models were involved in the Indian cases.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel