
If you buy or run AI security testing in India, read the new CERT-In document as a scoping instrument, not as a governance memo. The AI Security Auditing Guidelines v1.0 (cover dated 4 November 2026) connect how deeply an AI system must be tested to a four-level risk rating, and they name the design choices that push a system up a level. The practical result is a test catalogue and a re-test trigger that a penetration test report can be mapped against.
This post reads the document as a tester would. It also says plainly what we could not confirm, because the guidelines are long and parts of the text could not be read verbatim.
TL;DR
- The guidelines define four risk tiers (Low, Moderate, High and Critical) and tie testing depth to them: a baseline attack-surface assessment at Tier 1, threat modelling and API testing from Tier 2, comprehensive AI vulnerability assessment and penetration testing (VAPT) at Tier 3, and periodic red teaming at Tier 4.
- A section on advanced AI lists twelve escalation factors, including privileged access, persistent memory, delegation to other agents or tools, and no reliable way to roll back. Agentic and retrieval-augmented systems will often meet several of them.
- AI-specific testing must be repeated after material changes to the model, system prompt, retrieval corpus, embedding model, tool set, agent permissions, architecture, dependencies or provider model version.
- We found no statement of legal force or effective date in the parts we read, so treat "mandatory" claims in the press as unverified.
What the document is
The guidelines were prepared by a CERT-In sub-group working on Indian standards for cybersecurity auditing. They apply to AI systems developed, procured, deployed or operated by organisations in India, and to systems that process Indian data or serve Indian users. The intended audience includes auditors, CERT-In empanelled auditing organisations and sector regulators.
Two limits matter. First, the document uses "shall" for many individual controls, but we found no explicit sentence in the verbatim parts saying the document as a whole is binding or advisory. Second, no effective date appeared in the parts we read. Both points should be checked against the full PDF before anyone makes a compliance claim.
How testing depth scales by tier
The cybersecurity table in the guidelines (Table 24) is the most useful single page for a buyer. In summary:
- Tier 1: standard cybersecurity controls and a baseline AI attack-surface assessment.
- Tier 2: AI-specific threat modelling, API security testing, model integrity verification, AI logging and monitoring, and periodic AI security assessment.
- Tier 3: comprehensive AI VAPT covering all applicable test categories in the document's Section 10.4, plus adversarial testing, supply-chain scanning and tested rollback or fallback.
- Tier 4: everything in Tier 3, plus continuous monitoring, advanced adversarial testing, periodic red-team exercises, AI incident playbooks and forensic readiness. Elsewhere the text specifies annual red team exercises for Tier 4 systems.
The document describes the factors used to assign a tier: impact severity, autonomy of decision-making, data sensitivity, scale and exposure, reversibility, adversarial exposure and others. The detailed tier definitions and any scoring model sit in sections we could not read verbatim, so we do not reproduce them here.
The test categories
Section 10.4 lists the AI-specific VAPT categories. They are traditional infrastructure VAPT, API security testing, adversarial robustness testing, model extraction testing, retrieval-augmented generation (RAG) security testing, agentic AI security testing, agent memory testing, multimodal security testing, data poisoning detection, prompt injection testing for large language models, supply-chain scanning, and AI privacy attack testing (model inversion, membership inference, attribute inference and training-data extraction). The document describes the named tools as illustrative and allows equivalents that give adequate coverage, evidence quality and reproducibility.
For anyone who has written an AI test plan from the OWASP lists, this will look familiar. The OWASP Top 10 for LLM Applications (2025 edition) covers prompt injection, supply chain, data and model poisoning, excessive agency and vector and embedding weaknesses. The OWASP Top 10 for Agentic Applications for 2026 adds categories such as agent goal hijack, identity and privilege abuse, and memory and context poisoning. Testers will usually want to cross-reference all three lists when building a plan.
Why agentic features raise your tier
Section 9.5 covers generative, RAG, agentic, multimodal, multi-agent and autonomous systems. It lists twelve factors that, where they materially raise potential impact or adversarial exposure, require the assigned tier to be escalated with the rationale documented. Five of them are worth memorising:
- external or unrestricted public exposure;
- privileged access to production systems, data stores or administrative functions;
- persistent memory that can influence future actions;
- delegation to other agents or dynamically selected tools;
- inability to reliably interrupt, roll back or put the system in a safe state.
A customer-facing assistant with tool access, long-term memory and write permissions can satisfy several of these on day one, which in our reading pushes it toward a higher tier and fuller AI VAPT. The document also says input sanitisation alone must not be the primary control against prompt injection, which tells testers what kind of finding is expected: not "a filter was bypassed" but "the design relied on a filter".
Re-testing is event-driven
Section 10.4 requires AI-specific security testing to be repeated after material changes to the model, system prompt, RAG corpus, embedding model, tool or plugin set, agent permissions, deployment architecture, major dependencies or the provider's model version. That last item matters for teams that call a hosted model: a provider model version change is named as a trigger, so track which version you are running.
For delivery teams this is the sharpest operational consequence. A once-a-year pentest is hard to square with an event-driven rule if prompts and tools change weekly. Plan for a lighter regression test tied to the release pipeline, and keep a record of what changed and which tests re-ran. Our earlier notes on repository agent and editor configuration as an execution surface and on trusted publishing as an identity boundary describe two places where the change surface is wider than most teams track.
Where the guidelines meet rules already in force
Two existing frameworks line up with the document.
The 2022 CERT-In Directions under section 70B require reporting of listed cyber incidents within six hours of noticing them and retention of ICT logs for a rolling 180 days within Indian jurisdiction. Annexure I of that notice includes attacks affecting systems related to artificial intelligence and machine learning. The new guidelines repeat the six-hour and 180-day figures in their incident and logging tables, so an AI system's logging design is already a compliance question, not only a security one.
On the identity side, the US National Institute of Standards and Technology (NIST) published ten design tenets for agent identity management systems on 27 October 2026. It calls them an initial set and a starting point, not a final standard. Several of them read like test criteria: least and zero standing privilege, short-lived and revocable authorisation, architectural separation of policy from reasoning, and deterministic access decisions. Platform vendors are shipping related controls; for example, Microsoft's November 2026 Entra update lists a public-preview lifecycle policy for agent identities whose default inactivity policy notifies owners after 60 days and then disables the identity if no new activity occurs.
The CERT-In document also cites ISO/IEC 42001, NIST's AI Risk Management Framework, MITRE ATLAS and OWASP guidance as references. The NIST AI RMF 1.0 states that policies and resources should be prioritised according to the assessed risk level and potential impact of a system, which is the same logic the tier model applies. NIST's Generative AI Profile (AI 600-1, July 2024) is a companion for generative systems.
What to do with this now
- Tier your AI systems yourself before an auditor does, and write down which escalation factors apply and why.
- Ask your testers to map their report to the Section 10.4 categories and to state which were out of scope, so a missing category appears as a documented limitation rather than a silent pass.
- Put the Section 10.4 change list into your release process as a re-test trigger.
- Check that your logs for AI components meet the 180-day rule and that you can meet the six-hour reporting window for an AI-related incident.
The PentStark view
This is opinion. We think the most useful thing about the guidelines is that they give buyers a vocabulary for asking what "tested" means. A report that says "we tried prompt injection" is hard to compare with one that states the tier, the escalation factors considered, the categories covered and the categories excluded. We expect scoping conversations to move toward that structure whether or not a given organisation is formally bound by the document.
We would also be cautious about reading tool names in the tables as requirements. The text calls them illustrative, and a tier-based catalogue is only as good as the judgement used to assign the tier. Our methodology is described on the PentStark methodology page, but we make no claim here about how any specific engagement would map to these tiers.
What we could not confirm
We read the cover, executive summary, definitions, and the sections covering advanced AI controls and cybersecurity testing verbatim. Large parts of the document, including the full tier definitions, audit intensity per tier, auditor independence and empanelment rules, and the reporting format, were not available to us verbatim, so we have not described them. We also found no CERT-In notice stating whether the guidelines are mandatory. Check the full PDF and any official CERT-In communication before relying on this post for a compliance decision.
References
- CERT-In, AI Security Auditing Guidelines v1.0, 4 November 2026
- CERT-In Directions under section 70B(6) of the IT Act, 28 April 2022
- NIST, Design Tenets for Agent Identity Management Systems, 27 October 2026
- Microsoft Tech Community, What's new in Microsoft Entra: November 2026
- OWASP Top 10 for Agentic Applications for 2026
- OWASP Top 10 for LLM Applications, 2025 edition
- NIST AI Risk Management Framework 1.0 (AI 100-1)
- NIST AI 600-1, Generative AI Profile
Last verified: 2026-11-08
