THE ESSENTIALS IN 30 SECONDS

  • Systems are scored from 1 to 4 on eight risk factors; the total (8 to 32) places them in Tier 1 (low) to Tier 4 (critical), and some features such as autonomous code execution can push a system higher.
  • Higher tiers require independent audits by CERT-In-empanelled organisations; Tier 4 adds continuous monitoring, and sole-auditor engagements are not allowed for Tiers 3 and 4.
  • The framework covers AI developed, deployed or used in India or serving Indian users; according to The420.in it does not by itself create a new AI law.

India's national cybersecurity agency, CERT-In, has released a common framework for auditing artificial intelligence systems. The Artificial Intelligence Security Auditing Guidelines, version 1.0, are dated 4 November 2026 and were drafted by a CERT-In working group. Digit and The420.in describe the document as running to 107 pages.

How systems are classified

The guidelines apply to AI that organisations in India develop, buy, deploy or operate, and to systems that process Indian data or serve Indian users. Academic research projects, purely rule-based automation and personal consumer use are excluded.

Each system is scored from 1 to 4 on eight factors, among them impact severity, autonomy, data sensitivity, reversibility and exposure to attackers. The total, between 8 and 32, sets the tier: 8–12 is Tier 1 (low), 13–19 Tier 2 (moderate), 20–26 Tier 3 (high) and 27–32 Tier 4 (critical). Certain features override the sum, such as the ability to autonomously run code, make transactions or send external communications.

What audits look like

Audit depth rises with the tier. Tier 1 systems need a self-assessment or basic internal review each year. Tier 2 adds a documented internal audit and an independent audit by a CERT-In-empanelled organisation at least once a year. Tier 3 requires independent third-party audits, which Digit says may be needed as often as every six months when risk or material changes justify it. Tier 4 adds continuous monitoring. A single auditor cannot handle Tier 3 or 4 engagements, and small firms and startups get proportionality allowances only for the two lower tiers.

Audits cover nine areas: governance, data quality and privacy, model risk, fairness, transparency, security and incident reporting, monitoring and drift, human oversight, and third-party supply chains. The security controls repeat the six-hour deadline for reporting incidents and 180-day log retention. According to The420.in, both obligations come from CERT-In directions issued in April 2022.

Agents and legal status

Separate provisions cover generative AI, retrieval-augmented systems, multimodal models and AI agents. The document lists excessive agency, unsafe tool use, persistent memory and communication between agents among the risks, and asks auditors to test approval and override mechanisms separately.

The text uses mandatory wording for each tier's requirements, but how far it is legally binding is not spelled out in the published coverage. The420.in reports that the framework does not create a standalone AI law and separates existing legal duties from audit expectations.