CERT-In released the Artificial Intelligence (AI) Security Auditing Guidelines, version 1.0, dated 4 November 2026 — a framework for secure, ethical and auditable AI systems.
AI systems are classed into four risk tiers — Low, Moderate, High and Critical — and the tier decides audit depth, frequency, independence and the level of human oversight.
The framework sets nine core audit domains and seven guiding principles aligned with MeitY's India AI Governance Guidelines, with extra rules for generative, RAG, agentic and multi-agent AI.
AI incidents meeting CERT-In's reportable criteria must be reported within 6 hours, and logs kept for at least 180 days — the same limits as CERT-In's 2022 Directions.
Conventional AI answers a prompt. Agentic AI breaks a goal into steps, calls tools and APIs, runs code and delegates to sub-agents. The guidelines test agents for least privilege, credential leaks, memory poisoning, unsafe delegation and a working kill switch.
Simple Analogy: A calculator answers; an agent is an employee with the office keys.
Designates CERT-In as the national agency for incident response; the source of its power to issue guidelines and directions.
Set the 6-hour incident-reporting window and 180-day log retention that the AI guidelines extend to AI systems.
The seven CERT-In principles are aligned with MeitY's AI governance principles; the guidelines carry a section on integrating with them.
ISO/IEC 42001, ISO/IEC 23894, ISO/IEC 27001, NIST AI RMF, MITRE ATLAS and OWASP guidance for ML and LLM applications.
National nodal agency for cyber-incident response under Section 70B of the IT Act; operational since January 2004; empanels security auditing organisations
Parent ministry of CERT-In; author of the India AI Governance Guidelines and the IndiaAI Mission
GS Paper III > Cyber security; Science and Technology (AI)
General Awareness > Cyber security and regulators
General Awareness > Science and Technology
AI that plans multi-step tasks, uses tools and APIs and can delegate to sub-agents with little human input.
Hidden or malicious instructions that make an AI model ignore its rules or leak data.
Loss of accuracy as live data moves away from the training data, requiring continuous monitoring.
A human who must approve or can override an AI decision; its required level rises with the risk tier.