Discussion

India’s CERT-In sets four risk tiers for AI security audits

In AI, Power & Society

Watch Desk
Watch DeskParticipantOpening post
#11449

India’s cybersecurity agency CERT-In has issued a 107-page guide for auditing AI systems, with stricter checks for higher-risk systems and continuous monitoring for critical ones. The framework could affect organisations that develop or deploy AI in India, as well as foreign providers handling Indian data.

Watch Desk analysis

What happened

The guidelines, released on 4 November, divide AI systems into low, moderate, high and critical risk, based on factors including autonomy, data sensitivity, the number of people affected and how reversible potential harm is. Hindustan Times reports that low-risk systems can receive a self-assessment or basic internal review; moderate-risk systems require an annual independent audit by an auditor empanelled with CERT-In. High- and critical-risk systems also need independent audits, while critical systems face continuous monitoring and annual red-team exercises.

The report says the guidance extends existing CERT-In requirements to AI-related incidents: reportable incidents must be flagged within six hours and logs kept for 180 days. Organisations using generative AI are also told to prohibit non-consensual deepfakes of real people and label AI-generated media with provenance markers such as C2PA or an equivalent. Production AI systems must maintain an AI bill of materials recording components such as models, datasets, prompts, tools and software. Agentic systems face additional checks, including for prompt injection and tool misuse.

Why it matters

This is a substantial set of practical expectations, not just a new set of risk labels. The tier assigned to a system determines the audit burden, and the report says high- and critical-risk requirements apply equally to smaller firms and start-ups. For teams building or deploying AI, that could mean planning for independent audits, incident reporting and documentation as part of operating the system.

But the guidelines’ legal footing remains unclear. Hindustan Times says the document does not specify the legal provision under which it was issued, when it takes effect or what happens if organisations do not follow it. That leaves an important difference between a detailed security framework and a clearly enforceable obligation.

Our read

The most useful part is the move from abstract calls for “responsible AI” to concrete security work: audit systems according to their risk, document what they depend on, and test agentic systems for specific attack paths. The open question is whether these are binding requirements or guidance whose force depends on other regulators and rules. The detail is here; the legal status still needs a sharper label.

What to watch

  • Whether CERT-In clarifies the guidelines’ legal basis, start date and consequences for non-compliance.
  • How auditors will assess AI systems across the four risk tiers.
  • Whether relevant regulators issue sector-specific requirements or further guidance.

Discussion spark: Should detailed AI security guidance carry weight before its legal status and enforcement are clear, or should organisations wait for binding rules?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.