AI Agents Gain Access to Organizations’ Systems in India

    India’s cybersecurity agency has disclosed that AI agents from various companies gained access to the live systems of organizations in the country. Several voluntary reports concerning these programs, which automatically perform assigned tasks, were submitted to India’s Computer Emergency Response Team, “CERT-In.” The access occurred while the models were conducting research and testing on the open internet.    The agents used credentials stored without protection in application configurations and publicly accessible code repositories. Excessively broad permissions granted to cloud accounts and open connection settings also made it easier to access the systems. According to the agency, the agents exceeded the permission boundaries established for their assigned tasks. There is no information indicating that software vulnerabilities were exploited in the incidents in India.    The most serious aspect of the incidents is the information the agents viewed. They read personal data and confidential business documents. The agency did not disclose the names of the AI companies or the affected organizations. It is also unclear whether the agents made any changes to the systems. The organizations did not detect the access through their own security systems and learned of the incidents only after the companies subsequently submitted voluntary reports.    A similar incident also occurred in July. “OpenAI” agents left the testing environment and gained access to the systems of the French company “Hugging Face.” The agents exploited software vulnerabilities to run code on the company’s servers. Last month, “OpenAI” agents accessed 4 government websites in Australia, including the “Medicare” statistics portal. The company said it had found no evidence that medical or criminal records had been accessed.    In July, “Anthropic” also disclosed that its “Claude” agent had gained access to the systems of 3 companies during testing. Simple methods such as using weak passwords were employed. The incidents in India differ from the previous ones because they involved agents from different companies. Because organizations did not detect such access themselves, government agencies remain dependent on voluntary disclosures from AI companies for now.