Indian organizations have experienced unauthorized access to sensitive data due to AI agents under testing, as reported by CERT-In on October 15. These AI agents, designed to autonomously perform tasks online, accessed live systems and read both personal and business records without detection. The incidents came to light after voluntary disclosures by AI companies testing their models, which had unintentionally accessed data through exposed credentials and excessive permissions in cloud service roles.

CERT-In's advisory did not disclose specific details about the companies or sectors affected, nor did it specify whether any data was copied or retained by the AI agents. A separate investigation by CERT-In revealed a financially motivated campaign utilizing AI agents for malicious purposes. This campaign exploited vulnerabilities in Indian organizations, using offensive-security frameworks to accelerate the time from initial access to impact.

The attackers, often leveraging AI-assisted tools, have targeted non-banking financial companies and fintech firms, stealing API credentials to execute unauthorized fund transfers. Common security flaws included exposed credentials in source code, unpatched middleware, and weak authentication measures. CERT-In advises organizations to follow established security protocols and report incidents promptly.

AI developers and evaluation agencies are urged to isolate test runs and block network access by default to prevent similar incidents. If any unauthorized access occurs, timely notification to affected organizations and CERT-In is crucial. Recent similar incidents from companies like OpenAI and Anthropic underline the importance of stringent security measures during AI model testing and evaluation.