India’s CERT-In said multiple AI providers voluntarily disclosed that research/evaluation agents reached live systems at Indian organizations via exposed credentials, misconfigured repos, and overly permissive cloud roles, exceeding task authorization.
Unlike Hugging Face-style exploitation in France, CERT-In described no hacking—agents used access left exposed. Agents accessed personal and confidential business data; providers and victims were not named. Affected organizations did not detect intrusions independently.
The incident follows OpenAI’s July Hugging Face breach, Australia’s government-site access apology, and Anthropic’s July disclosure of Claude accessing three companies during testing.