CERT-In's 107-Page AI Audit Rules Set Four Risk Tiers

CERT-In, India's national computer emergency response team, published its 107-page Artificial Intelligence Security Auditing Guidelines on Nov. 4, 2026. They apply to AI systems that organisations in India develop, buy, deploy or operate, and to systems that process Indian data or serve Indian users, according to MediaNama and The420.in.

The guidelines are not a new statute. The420.in quotes them as saying the framework "does not, by itself, establish a new standalone AI law." What they do is give auditors, and the regulators who rely on them, a common checklist. The document is separate from the payment-API advisories we covered in CERT-In's naming of the CRAFTY FALCON actor, though both arrived in the same week.

How the four tiers work

A system is classed Low, Moderate, High or Critical on factors including impact severity, decision autonomy, data sensitivity, scale and exposure, reversibility and sector criticality. MediaNama's summary of the tiers:

Tier	Description	Audit requirement
Low	Advisory or assistive, limited autonomy	Streamlined requirements
Moderate	Influences business processes, defined human oversight	Annual independent audit by CERT-In-empanelled auditors
High	Significant effects on operations, users or sensitive data	Periodic independent audits, possibly every six months
Critical	Severe or systemic risk, including critical infrastructure	Empanelled audits and continuous monitoring

Tiers can be raised if a system's use, impact or exposure changes. The420.in adds that systems must be reassessed when capabilities or operating environment change.

Remediation clocks by finding severity

Auditors rate findings Critical, High, Moderate or Low. MediaNama reports these deadlines:

Finding	Remediation	Corrective action plan	Progress reports
Critical and High	Not specified	Required	Monthly
Moderate	Plan within 30 business days, implement within 60 days	Required	Quarterly
Low	Within three months	Not specified	Annually

A re-audit is required when an audit opinion is unsatisfactory or when two or more Critical findings arise within three months. A regulator can demand one sooner.

Extra scrutiny for agents, RAG and generative models

For generative AI the guidelines name prompt injection, sensitive-information disclosure, data and model poisoning, hallucinations and improper output handling. Retrieval-augmented generation systems get a review of retrieval pipelines, vector databases and embeddings. Agentic systems are examined for tool use, excessive agency, least-privilege access and persistent memory. Deepfakes require explicit consent in relevant contexts, with measures such as C2PA content provenance.

Auditors may use AI tools for evidence analysis and report drafting, but only with approved scope, confidentiality controls and human review. The guidelines say, as MediaNama and The420.in each quote them with slightly different wording, that AI-generated findings cannot replace sufficient evidence or independent human judgement. That sentence applies equally to tools that audit coding-agent sessions.

What is old and what is new

Incident reporting is the figure most likely to be misread. Reportable AI security incidents must be notified to CERT-In within six hours and logs kept for 180 days. The420.in says this is not a new obligation: it comes from CERT-In's directions of April 28, 2022, which the guidelines fold into the AI framework. An inaccurate chatbot answer does not by itself count as a reportable incident.

RBI's separate deadlines for bank APIs

MediaNama reported on Nov. 6, drawing on an exclusive by Arti Singh of The Head and Tale, that the Reserve Bank of India met bankers for nearly two hours and sent a confidential advisory dated Nov. 3. As reported, banks must hold at least one safeguard on API fund transfers, with existing integrations compliant by Dec. 3, 2026 and new ones before going live. NBFC API clients must complete CERT-In audits by Feb. 3, 2027, and audits must be repeated yearly. Per-transaction and daily limits take effect immediately, and intra-day velocity limits within one month. The advisory text itself was not available for this report, so these dates rest on MediaNama's account.

The next date to watch is Dec. 3, 2026, when the first RBI API deadline falls.