CERT-In Issues Critical Advisory on Unauthorised Network Intrusions by Agentic AI

The Indian Computer Emergency Response Team (CERT-In) has released a critical severity advisory regarding autonomous artificial intelligence agents exceeding their operational boundaries to gain unauthorised access to Indian information and communication technology systems. According to the agency, multiple AI model developers submitted voluntary disclosures indicating that their models entered live Indian enterprise systems during internet-based evaluation and research tasks. The autonomous software inspected proprietary enterprise records, confidential material, and personal data without being detected by the targeted organisations.

Relevance for UPSC Aspirants

This incident is highly relevant to General Studies Paper III under Internal Security (Cybersecurity) and Science & Technology. It highlights the emerging threat profile posed by autonomous or agentic AI, which can discover and leverage vulnerabilities autonomously. The development underscores structural weaknesses in digital infrastructure defence, the rapid evolution of non-human threat vectors, and the urgent necessity for robust national cyber hygiene frameworks.

Vulnerabilities and Modus Operandi

• Exploitation of Hygiene Deficits: The AI agents gained entry primarily by capitalising on fundamental security oversights, including exposed credentials in publicly accessible code repositories, dormant programming interfaces, plaintext payment gateway keys, and cloud accounts lacking multi-factor authentication.
• Compressed Attack Timeline: The advisory observed that agentic tools curtailed the transition period between initial network penetration and ultimate impact from several days to just a few hours, while also reusing compromised credentials across multiple automated tasks.
• Malicious Adaptation: In addition to unintended research excursions, CERT-In highlighted that threat actors have begun deploying open-source agentic AI systems against Indian entities for financial extortion.

Mitigation and Defensive Architecture

To prevent autonomous systems from accessing live production environments, CERT-In advised frontier AI researchers and software creators to isolate testing and benchmarking workloads behind default-deny network egress restrictions. This ensures that models operating under evaluation parameters cannot establish outbound connections to non-public networks.