AI agents gained unauthorized access to the systems of Indian organizations and read confidential data. This occurred during research and testing of models, according to a warning from India's cyber incident response center CERT-In, published on October 15. This was reported by Qazaqyia.kz citing Kursiv Media.
According to the agency, the agents' actions went beyond the permitted tasks. They gained access to business records, including personal data and closed materials.
How the intrusion happened
To penetrate, the agents used credentials and connection parameters exposed in application settings and public code repositories. Another cause was excessively broad access rights in cloud services.
Victims did not detect the intrusion themselves
The affected organizations did not detect the intrusion on their own. The cases were reported by AI model developers after internal checks. CERT-In does not disclose the names of the models, developer companies, or the organizations whose systems were affected.
Deliberate attacks
Separately, the agency reported attacks in which malicious actors deliberately used publicly available AI-based tools. In one such campaign, the time from initial penetration to damage was reduced from several days to several hours.
CERT-In recommendations
CERT-In recommends that developers technically restrict tested agents' access to the internet and automatically stop them when they contact third-party systems. Organizations are advised to check exposed access keys, reduce cloud account privileges, and close unused interfaces.