CERT-In and AI in 2026, what changed and what to do.

Two new documents, one old rule

CERT-In, India’s national agency for cyber incidents, issued two documents about AI in quick succession in 2026:

• The blueprint, 25 May 2026. Its full title is the Blueprint for Reducing Exposure and Defending against AI-Assisted Vulnerabilities Exploitation in Digital Infrastructure. It is written for organisations of every kind, not only AI companies, because the threat it describes, attackers using AI to find and exploit weaknesses faster, affects everyone.
• The vendor guidelines, 10 June 2026. These are for technology providers: software and hardware makers, cloud and managed service providers, system integrators and digital service providers, whether they are based in India or abroad. They build on a CERT-In advisory of 26 April 2026 on cyber risks from frontier AI.

Neither replaces CERT-In’s Directions of April 2022, which remain binding: report listed cyber incidents within six hours, keep 180 days of logs inside India, and sync your clocks to official time servers. Our CERT-In guide covers those rules in full.

Why CERT-In moved now

The message running through both documents is speed. AI tools now help attackers read code, find weaknesses and write working exploits much faster than before, so the gap between a flaw becoming known and it being used in attacks keeps shrinking. A patch cycle that waits for the monthly maintenance window leaves the door open for weeks.

CERT-In’s answer has three parts: fix the most dangerous flaws in hours rather than weeks; know exactly what software and AI you run, so you can tell at once whether a new flaw affects you; and keep a firm grip on AI inside your own organisation, especially AI agents that can act on their own.

The new patch clocks

The blueprint sets risk-based deadlines to patch, mitigate or remove the exposure. The tightest, 12 hours, is for flaws already being exploited in the wild on systems reachable from the internet, “where feasible”. That phrase matters: the blueprint accepts that not every system can be patched that fast, and suggests interim measures instead, such as isolating the system, restricting access, putting web application or API protection in front of it, or watching it more closely.

Few organisations can meet a 12-hour clock by hand. It needs three things in place before the flaw appears: a complete list of internet-facing assets, a way to tell within minutes which of them are affected, and pre-approved emergency change rules, so nobody waits for the next change meeting. Our vulnerability management service is built around exactly these clocks.

What it asks of the AI inside your company

The blueprint treats AI as both a threat and something to protect. Its main asks:

• An AI inventory. Know which AI systems, models and AI-powered tools you run, who owns each, and what data each one touches.
• Shadow AI under control. Watch for AI tools staff use without approval, publish an AI usage policy, approve new AI integrations before they go live, and restrict uploads of sensitive information to public AI platforms. See shadow AI discovery.
• An AI bill of materials (AIBOM) next to software, cryptography and quantum bills of materials (SBOM, CBOM and QBOM). When a model, library or dataset turns out to be flawed or tampered with, you can tell in minutes whether you use it.
• Limits for AI agents. For agents that act without continuous human supervision: defined boundaries and permissions, continuous monitoring, audit logs and an emergency shutdown mechanism. Our MCP security guide shows why each of these matters.
• Deepfake defences. Procedures to verify that a call, message or video from a senior leader is genuine, monitoring for impersonation, and a clear escalation path, including to law enforcement. Our deepfake fraud drills test exactly this.
• Awareness. Training on AI-written phishing, deepfake impersonation and social engineering.

The blueprint also announced a new CERT-In AI Cyber Defence Center. Incident reporting does not change: cyber incidents still go to CERT-In within six hours.

The 60-day plan

The blueprint ends with a phased roadmap. The first week closes the doors attackers use most: multi-factor login, scanning, patching known exploited flaws and being ready to report. The next three weeks build visibility and governance, including the AI inventory. The final month is about testing: red team exercises, adversarial simulations and adversarial testing of AI systems, including checks that models have not been tampered with. That last phase is where LLM red teaming and red team operations come in.

The June rules for technology vendors

If you build or sell technology that Indian organisations use, the guidelines of 10 June 2026 ask more of you:

• Disclose at once. Report critical (CVSS 9.0 to 10.0) and high (7.0 to 8.9) vulnerabilities to CERT-In and to every affected Indian customer as soon as they are found or confirmed, with interim mitigations and a timeline for the fix.
• Fix AI-exploitable flaws on a clock, as in the table below.
• Keep live bills of materials for software, hardware, cryptography and AI (SBOM, HBOM, CBOM and AIBOM), keep them current, and share them with customers and CERT-In.
• Show your testing. Keep evidence that security assessments ran, including AI-assisted code analysis and automated vulnerability discovery, and include breach and attack simulation results in an assurance report.

Large Indian customers can be expected to copy these terms into contracts and security questionnaires. Vendors that can already hand over a current AIBOM and a recent test report will get through those reviews faster.

What is binding, and what is advice

A ten-point checklist

1. List every internet-facing system, and who can patch each one out of hours.
2. Track known exploited vulnerability feeds and match them to your systems automatically.
3. Agree emergency change rules now, so a 12-hour fix needs no meeting.
4. Decide interim measures (isolate, restrict, filter, watch) for systems you cannot patch fast.
5. Build an AI inventory: every AI system, model, AI-powered SaaS tool and its owner.
6. Look for shadow AI, and publish a short AI usage policy.
7. Start an AIBOM next to your SBOM, beginning with your customer-facing AI.
8. Give every AI agent written limits, logged actions and a tested off switch.
9. Set up a way to verify urgent requests from executives, and drill it against deepfakes.
10. Put adversarial testing of your AI systems on the calendar, and keep the evidence.

Where we fit

We help organisations act on the blueprint end to end: finding AI and shadow AI, setting up patch clocks, testing AI systems and agents, and building the evidence. We are not a CERT-In empanelled auditor. Where you need a formal audit, we prepare you and work alongside the empanelled auditor you choose. See CERT-In readiness.