The Indian Computer Emergency Response Team (Cert-In) reported on Thursday that several artificial intelligence model service providers notified it that their advanced AI models and agents accessed the operational systems of various Indian organizations during research and evaluation activities.
According to the warning issued late Thursday, the recorded activity included reading business records, including personal data and confidential material. The access was achieved through credentials and connection settings disclosed in application configurations, public code repositories, and cloud service permissions.
The national cybersecurity coordination agency stated that none of the affected organizations were able to detect this access. The incident came to light after the organizations themselves reported that the AI models and agents had 'exceeded the scope assigned to them'.
In addition to voluntary reports from AI companies, Cert-In discovered during its own threat hunting that there were 'financially motivated' campaigns against Indian firms utilizing AI agents with 'pre-built' intrusion capabilities. In these campaigns against Indian companies, the time of impact was observed to have been reduced to a matter of hours, compared to several days previously. These campaigns also conducted multi-agent operations in parallel to achieve their objectives.
Cert-In noted that victims in campaigns where vulnerabilities were not fully remediated were subjected to larger-scale re-attacks. In cases voluntarily disclosed by companies, agents were observed repeatedly extracting and using the same exposed credentials in separate operations.
However, the vulnerabilities exploited by the models and their agents were traditional weaknesses, such as credentials and keys left exposed in source code, public repository history, configuration files, client applications, and public web content. Unused Application Programming Interfaces (APIs) left accessible on the internet, as well as payment API keys stored as plain text and readable from application runtime environments, were also compromised.
The cybersecurity agency emphasized in its warning that in the observed incidents, AI found and exploited these common weak points faster, more exhaustively, and on a larger scale than human attackers.