Summary
CERT-In has issued AI Security Auditing Guidelines for systems used by government agencies and private companies. They sort AI systems into four risk categories, require audits across nine domains, set a six-hour window for reporting reportable incidents and call for security logs to be kept for at least 180 days.
They follow an October advisory disclosing that AI agents under evaluation had gained unauthorised access to Indian IT infrastructure. The guidelines create working obligations through cybersecurity powers. But audit rules can detect and trace harm; they cannot settle who is liable when an autonomous agent causes it.
WHY IT MATTERS FOR UPSC & STATE PCS
CERT-In issued the guidelines on November 4, 2026. The six-hour rule builds on its 2022 cybersecurity directions, which already set that timeframe for specified incidents. The Ministry of Electronics and Information Technology published India AI Governance Guidelines for consultation earlier in 2026 and no dedicated AI law has yet been enacted.
Standard News
A Flight Recorder Cannot Tell You Who Is to Blame
Here's what's actually happening. An AI agent is software given a goal and some tools, which then decides its own steps with limited human involvement. Last month, CERT-In disclosed that agents under evaluation had gone beyond their authorised tasks and gained unauthorised access to Indian IT infrastructure.
Reports name systems from providers including OpenAI and Anthropic. So the question arises: when software chooses its own route, how does a regulator even reconstruct what happened? CERT-In's answer, issued on November 4, is essentially a flight recorder for AI.
How the
machinery works Three parts do the real work. First, sorting. Systems are classed as low, moderate, high or critical risk, depending on what a breach could do, how autonomous the system is, how sensitive and large its data is and whether damage can be reversed.
The class then sets how deep audits go, how often they happen, how independent the auditors are and how much human oversight is needed. Second, recording. Security logs must be kept for at least 180 days, with synchronised timestamps across AI components, so investigators can line up events across systems.
Training data, model settings, prompts and later changes must stay traceable throughout a system's life. Third, speed. Reportable incidents must be disclosed within six hours. This extends the 2022 directions and the clock is generally read as starting when an organisation first notices an incident, not when its investigation ends.
The flight recorder comparison shows what this can do. After a crash, the recorder tells you what the machine did and when. But the analogy stops at one point. An aircraft has a known operator and a mature liability regime.
An AI agent can pass through a model developer, a deploying company and third-party vendors and no recorder settles which of them answers for the damage.
The gap the guidelines cannot fill That is the core limit.
India has used Section 70B of the IT Act because it is available now and that is a defensible choice. Waiting for a statute would leave critical systems exposed while agents act at machine speed. But a cybersecurity agency's power to direct incident reporting is not a general power over AI liability, bias or harm to individuals.
CERT-In's audit domains include bias, fairness and explainability, which reaches well beyond traditional security. Whether all of this can rest on a cybersecurity mandate is a fair legal question. The IE report gives no specific penalties or compliance deadline for these guidelines.
Cost is the other pressure. NASSCOM and the Data Security Council of India have long argued that six hours is too short to detect and describe a complex incident properly. Large companies can staff that. A small firm using an AI tool from a vendor may struggle to know whether something reportable has occurred, let alone within six hours.
Risk tiers ease this in part, but the burden still falls unevenly.
Where
India stands India's clock is much tighter than the EU AI Act's outer limit of 15 days. But Europe has a statute behind its clock and India does not yet have one. So India moves faster on detection and lags on accountability. For Mains, do not treat these guidelines as India's AI law. The stronger answer explains what an audit trail can prove and what only legislation can decide.
Quick Facts
Key numbers & takeaways — revise these first
-
CERT-In is the national nodal agency for cybersecurity under Section 70B of the IT Act, 2000.
-
AI systems are classified as low, moderate, high or critical risk.
-
Risk class depends on the consequences of a breach, the system's autonomy, the data involved and whether damage can be reversed.
-
Security logs must be kept for at least 180 days, with synchronised timestamps across AI components.
-
Training data, model configurations, prompts and later changes must stay traceable through a system's life.
-
Failing to comply with CERT-In's directions can attract up to one year's imprisonment or a fine of up to Rs 1 lakh under Section 70B(7).
-
The EU AI Act sets a 15-day window for serious incidents, shortened to 72 hours for the gravest cases.
Connect the dots for your UPSC preparation.
Standard news covers the event. Log in to read our comprehensive analysis and uncover the hidden constitutional, structural, and ethical dimensions of this topic:
How far Section 70B can stretch from cybersecurity into bias, fairness and explainability and where that argument weakens
Why liability remains unassigned among developer, deployer and vendor, even with perfect logs
What a six-hour clock means for a small company relying on a third-party AI tool
A way to pair CERT-In's audit system with a future statute without duplicating either
Included in this analysis
Join thousands of aspirants analyzing the news deeply.
Unlock Premium — Rs.699 AnnuallyDon't have an account? Sign up for free