The first question a CISO or CIO will ask about the new CERT-In document is simple: what does it cost us, and when? The honest answer depends on one number you may not have yet, which is how many AI systems sit in your organisation and what tier each one falls into.
That is the pattern across the whole document. It does not ask for a new policy binder. It asks you to prove, with sampled evidence, that your AI systems are known, owned, tested, monitored and overseen by people who can actually intervene. For a CISO, that is a security and assurance problem. For a CIO, it reaches into model pipelines, platforms, vendors and recovery plans.
This blog covers what changes for each role, the eight points that deserve board attention, a 90-day plan, and the parts of the document that are still unclear.
What do the CERT-In AI Security Auditing Guidelines mean for CISOs and CIOs?
They move AI from an innovation topic to an audit topic. The CISO is accountable for AI security, incident reporting and assurance evidence. The CIO is accountable for the platforms, pipelines, vendors and recovery capability that the audit will inspect. Both sit under the AI governance structure that the tier demands.
The tier sets how heavy that structure is. Tier 1 needs a designated AI owner. Tier 2 needs an AI governance function. Tier 3 needs a formal AI Governance Committee that meets at least quarterly with minutes. Tier 4 needs board or senior-management oversight with defined independent assurance.
| Tier | Audit mechanism | Technical testing | Governance |
|---|---|---|---|
| 1 Low | Self-assessment or basic internal review | Basic security assessment | Designated AI owner |
| 2 Moderate | Documented internal audit; independent audit by an empanelled auditor at least annually per Section 5.4 | AI threat modelling, API and security testing | AI governance function |
| 3 High | Independent audit by an empanelled auditor; six-monthly where risk or change warrants | Full AI security assessment and VAPT | AI Governance Committee |
| 4 Critical | Independent audit with continuous monitoring | Advanced adversarial testing and red team exercises | Board or senior-management oversight |
What are the key takeaways for CISOs and CIOs?
Eight points matter most. The first four are about knowing and proving. The last four are about control.
01. Tiering decides your audit budget
Two Tier 3 systems cost far more to assure than ten Tier 1 systems. The tier sets audit frequency, the amount of technical testing, whether model validation must be independent of the development team, and what the governance structure looks like. Get the tiering right before you plan spend.
Scoring uses eight factors rated 1 to 4: impact severity, autonomy, data sensitivity, scale and exposure, reversibility, sector criticality, adversarial exposure, and complexity. A composite of 20 to 26 is Tier 3. A composite of 27 to 32 is Tier 4. Escalation triggers, such as privileged production access or actions that cannot be reversed, can lift a system above its score.
02. The inventory and the auditor’s tier will both be challenged
The auditor starts with your AI system inventory and re-scores the systems independently. Where the auditor’s tier differs from yours, the higher tier applies for scoping. A missing or incomplete inventory is a governance non-conformity.
Shadow AI is the practical risk. Think of a business unit that switched on an AI assistant inside a SaaS platform, or a developer team calling a hosted model API from a production service. If it processes company data and nobody listed it, it is a finding. The cheapest fix is a discovery exercise that combines network egress logs, procurement data and a business-unit questionnaire.
03. Independent audit capacity is the bottleneck
Tier 3 and Tier 4 systems need independent audits by CERT-In empanelled auditors. Audit teams must include at least one member with demonstrated AI and machine learning competence and one with an audit certification such as CISA, CIA or ISO 27001 Lead Auditor. Sole-auditor engagements are not allowed.
Few firms have both skill sets on one team today. If you expect Tier 3 systems, ask auditors now about team composition, availability and how they handle independence. An auditor who also designed your controls cannot assess them, so a consultant who helped you build the AI impact assessment may be unable to audit it in the same cycle.
Also ask how the auditor uses AI tools on your data. The Guidelines allow AI-assisted audit work only with documented approval, your written authorisation, data redaction, tool validation and competent human review, and they require the report to disclose it.
04. Human oversight will be tested, not accepted
Section 11 adds a separate test of contestability. Auditors inspect the criteria that trigger human review, sample AI decisions to check the review happened, and test whether approval gates can be bypassed through an alternate interface or API. They also look at reviewer training, override frequency, reviewer agreement rates and repeated exceptions.
A common failure: an approval step exists in the workflow tool, but the same action is reachable through a direct API call that skips it. The auditor will find that. For high-risk and critical systems, the auditor performs or observes practical tests of review, override, interruption and recovery.
05. The CIO owns the pipeline, and the pipeline is in scope
Training, evaluation and deployment pipelines count as part of the security boundary. The controls the Guidelines list include segregated development, test and production environments, authenticated pipeline identities with least privilege, signed or hashed production model artefacts verified before loading, evaluation gates before promotion, separate approval from the person who started the training job, and a tested rollback to a known-good model.
Automated retraining cannot push a new model into production without validation and approval. Tier 3 and 4 model changes should use staged, canary or shadow deployment where technically feasible. Production AI systems need an AI Bill of Materials covering model, datasets, dependencies, prompts, guardrails, vector stores, tools and external APIs.
06. Agentic AI has its own rules
Agents need documented identities, approved tools and defined permission boundaries. They run under least privilege and do not inherit permissions from users, tools or other agents. Consequential actions go through human approval gates. High-risk agentic systems must support emergency interruption, credential revocation and rollback.
Agents that connect to external tool servers or Model Context Protocol (MCP) services need an approved inventory of permitted connections, authenticated identities, validated tool descriptions and logged tool calls. New or dynamically discovered tools need a security assessment before they reach production. Multi-agent systems add agent identity, authenticated agent-to-agent messages, controls on recursive delegation, and the ability to suspend a whole workflow.
07. Vendor contracts are audit evidence
Third-party and supply chain assurance is one of the nine mandatory domains for Tier 2 to 4 systems. The auditor checks vendor due diligence, model and component provenance, contractual security obligations, provider change management, consumer-side monitoring of AI-as-a-Service, and exit and fallback plans for critical dependencies.
Contracts should define vulnerability notification, incident notification, change notification and audit and assurance rights. If a provider will not allow inspection, you need independent assurance reports or black-box testing instead, and the limitation goes into the audit record. “The vendor would not share” does not make the control not applicable.
08. Incident response and board reporting
CERT-In reportable AI incidents go to CERT-In within 6 hours, and logs are kept for 180 days with synchronised time across AI components. Neither requirement is new. They come from the 2022 Directions. What is new is the AI layer: SOC detection rules for model scraping, repeated probing and unexpected model version changes, plus playbooks for data poisoning, model extraction, prompt injection breaches and deepfake incidents.
For Tier 4, the Guidelines call for annual red team exercises simulating AI-specific attacks, with results fed back into SOC detection rules. Board reporting follows maturity: the five-level model scores each of the nine domains and takes the weakest as the overall level.
What should CISOs and CIOs do in the next 90 days?
Start with the inventory and the tier scores, then close the gaps that need the longest lead time. The list below runs in order.
- Weeks 1 to 4: build the inventory. Combine procurement data, network logs and business-unit input. Record purpose, owner, data categories, model type, users and deployment status for each system.
- Weeks 3 to 6: score and tier. Rate each system on the eight factors. Record the rationale and evidence for each score. Check the escalation triggers.
- Weeks 4 to 8: test your reporting path. Run a tabletop for an AI incident. Confirm the 6-hour route to CERT-In, log retention of 180 days and time synchronisation.
- Weeks 4 to 10: fix pipelines and agents. Close gaps in artefact signing, environment separation, approval gates, agent permissions and MCP connection inventories.
- Weeks 6 to 12: review vendors. Add audit, incident, vulnerability and change notification rights. Request AIBOM or SBOM data and exit plans for critical AI dependencies.
- Weeks 8 to 12: brief the board and book the auditor. Present the tier distribution and the gaps. For Tier 3 and 4 systems, agree audit windows and team composition early.
The split of work is usually as follows.
| Area | Typically led by | Evidence the audit will ask for |
|---|---|---|
| Inventory, tiering, AI impact assessments | CISO with risk and business owners | Inventory, scoring sheets, AIIA reports |
| AI threat model, VAPT, adversarial testing | CISO | Test reports with reproducible records |
| SOC integration and incident response | CISO | Detection rules, playbooks, tabletop records |
| MLOps and LLMOps controls, rollback | CIO | Pipeline logs, artefact hashes, rollback test results |
| AIBOM and model registry | CIO | Registry export, AIBOM per production system |
| Vendor and AIaaS assurance | CIO with procurement and legal | Due diligence files, contract clauses, exit plans |
| Human oversight and decision review | Business owner with CISO | Override logs, reviewer training, effectiveness tests |
What does the audit report look like?
The report has an executive summary of no more than two pages, scope and methodology, an overview of the AI system portfolio, detailed findings, a finding summary matrix, management responses, positive observations and the overall assurance opinion. Each finding states the condition, the criteria, the root cause, the effect and a recommendation, and the organisation responds with owner and date.
Severity drives the clock. Critical findings need immediate escalation and containment. Moderate findings need a plan within 30 business days and implementation within 60 days. An Unsatisfactory opinion or two Critical findings triggers a re-audit within three months. That makes the opinion a board matter, not a technical footnote.
Where are the CERT-In AI Security Auditing Guidelines still unclear?
Two points deserve a watch. The first is legal force. The text uses mandatory wording but names no penalty or deadline, and commentators note the framework does not by itself create a standalone AI law. Check whether your sector regulator adopts it by circular.
The second is Tier 2. One section requires an independent audit by an empanelled organisation at least annually, while another table lists a documented internal audit as the primary mechanism. The difference changes cost and timing for every Moderate system you own. Until CERT-In clarifies, budget for the stricter reading.
Summary
CISOs and CIOs should read the CERT-In AI Security Auditing Guidelines as a test of evidence. The tier sets the cost, the auditor can overrule your self-assessment, and independent audit capacity will be tight for higher-risk systems. Oversight, pipelines, agents and vendor contracts are where audits will probe hardest. Build the inventory first, tier it honestly, and brief the board before the auditor does.
Frequently Asked Questions
Build a complete AI system inventory and tier every system. The auditor starts there, and an incomplete inventory is a governance non-conformity. Everything else, from audit type to testing depth, follows from the tier.
Who is accountable for AI security under the Guidelines?
Every production AI system needs a clearly assigned owner, and governance scales with the tier. Tier 3 needs a formal AI Governance Committee, and Tier 4 needs board or senior-management oversight. The CISO usually holds security, incident reporting and assurance evidence, while the CIO holds platforms, pipelines and vendors.
Does the auditor accept our own risk tier?
No. The auditor validates or reassesses the tier using the eight-factor method and does not rely only on yours. If the two differ, the higher tier applies for scoping.
How should we prepare for an independent audit of a Tier 3 system?
Assemble the pre-audit evidence pack (governance records, model registry, validation reports, bias tests, threat model, VAPT, supply chain assessment, oversight test records), run an AI-specific penetration test, and confirm the audit team’s composition. The team needs at least two people, including one AI and machine learning specialist and one CISA, CIA or ISO 27001 Lead Auditor.
Are the 6-hour reporting and 180-day log rules new?
No. They come from CERT-In’s Directions of 28 April 2022. The Guidelines apply them to AI incidents and require time-synchronised logs across AI components for forensic readiness.
What happens if the audit finds serious problems?
Critical findings need immediate escalation and containment. An unsatisfactory opinion or two critical findings trigger findings, triggers a mandatory re-audit within three months covering all critical and high findings and their related domains.
Can we use generative AI to prepare for or support the audit?
Yes, with controls. Auditors and organisations may use approved AI tools for document review, mapping and drafting, but outputs are working aids, not evidence. Use needs authorisation, data protection, validation of the tool, traceability to original evidence and human review.
Does an ISO/IEC 42001 certification satisfy the Guidelines?
Not on its own. The Guidelines align closely with ISO/IEC 42001 and map their controls to its clauses and Annex A, so an ISO 42001 management system helps. The Guidelines still require tiered control testing, technical security testing and an independent opinion for higher-risk systems.




