NewsTree24 | Published November 6, 2026 | Original document date: November 4, 2026
India’s Computer Emergency Response Team (CERT-In) has published a 107-page set of Artificial Intelligence Security Auditing Guidelines, giving organisations a risk-based framework for assessing AI systems across development, deployment and operation. The document is relevant not only to Indian businesses: it says the framework covers organisations operating AI systems in India and systems processing Indian citizens’ data, including international AI-as-a-service and platform providers serving Indian customers.
The guidance arrives as companies deploy generative AI, retrieval-augmented generation (RAG) and autonomous agents into customer-facing and internal workflows. Its central message is that AI assurance must cover the data, model, tools, infrastructure and people around a system—not just the model’s final answers.
Four risk tiers determine audit depth
CERT-In’s framework classifies AI systems as Low, Moderate, High or Critical. The risk assessment considers eight factors: impact severity; autonomy and actionability; data sensitivity; scale and exposure; reversibility and recoverability; operational or sector criticality; adversarial exposure; and system complexity and auditability.
- Low risk: a self-assessment or basic internal review with documented evidence.
- Moderate risk: a documented internal audit, plus an independent audit by a CERT-In-empanelled organisation at least annually.
- High risk: periodic independent third-party audits, with more frequent review when material changes, incidents or risk conditions warrant it.
- Critical risk: an independent audit by a CERT-In-empanelled organisation and continuous monitoring.
The framework says the tier should reflect the system’s real capabilities and exposure. A customer-support assistant with no action permissions may present a different risk from an agent that can approve payments, change production systems or access sensitive records.
Nine audit domains cover the AI lifecycle
The guidelines organise assurance around nine areas: AI governance and accountability; data lifecycle, quality and privacy; model risk management; ethics, bias and fairness; explainability and transparency; security and adversarial resilience; continuous monitoring and drift management; human oversight; and third-party and AI supply-chain assurance.
That scope extends beyond model testing. It includes data provenance, prompt and guardrail configuration, API security, model integrity, change control, monitoring, vendor due diligence and the evidence needed to reconstruct security-relevant events.
Extra checks for generative AI and autonomous agents
The document includes dedicated requirements for generative AI, RAG, agentic AI, multimodal systems and multi-agent architectures. Among the risks it tells auditors to examine are prompt injection, sensitive-information disclosure, excessive agent permissions, unsafe tool use, model or data poisoning, vector-store weaknesses, persistent-memory risks and inter-agent communication.
For agentic systems, the framework emphasises identifiable service identities, least-privilege access, authenticated tool connections, traceable tool calls and human approval for consequential actions. High-risk agents should support interruption, credential revocation and recovery rather than being allowed to continue uncontrolled after an abnormal event.
Incident reporting and log retention
The security-audit section refers to AI-related incident reporting within the applicable CERT-In six-hour timeline and forensic readiness supported by 180-day log retention. These references should be read alongside existing CERT-In directions and other applicable obligations; they are not a reason to assume that every organisation has received a new, separate statutory deadline from this document.
Important implementation caveat
CERT-In’s official guideline page dates the document November 4, 2026. The PDF sets out tier-specific audit requirements, but it does not itself state a separate commencement date or penalties for non-compliance. Hindustan Times also noted that the document does not identify the legal provision under which it was issued or specify when it applies. Organisations should therefore map the guidance against existing CERT-In directions, sectoral rules, contracts and advice from qualified legal and compliance teams rather than assume that this PDF alone creates a new statutory deadline.
Practical steps for AI and technology leaders
- Inventory AI models, datasets, prompts, tools, APIs, vendors and production integrations.
- Assign an accountable business owner and classify each system using its actual autonomy, access and potential impact.
- Review agent permissions, service identities, network egress, approval gates and emergency-stop procedures.
- Test prompt injection, data leakage, unsafe tool use and third-party model dependencies before deployment and after material changes.
- Keep time-synchronised logs and evidence that can support incident investigation, audit and remediation.
For related NewsTree24 coverage, see our analysis of Snowflake’s enterprise AI governance controls, IBM’s agent identity release and our guide to AI cybersecurity tools for businesses.
Sources and attribution
- CERT-In: Artificial Intelligence (AI) Security Auditing Guidelines v1.0, dated November 4, 2026
- CERT-In official guidelines listing
- Business Standard: CERT-In calls for AI security audits, November 4, 2026
- Hindustan Times: CERT-In issues AI security audit guidelines, November 5, 2026
Featured image: Business analytics dashboard representing AI monitoring and audit evidence. Photo by Luke Chesser on Unsplash.