CERT-In's New AI Rules: 6-Hour Reporting, Mandatory Kill Switches

If an AI system your company runs goes wrong in India starting this month, the clock starts ticking the moment you notice it. Not next week, not after an internal review — six hours, full stop, or you're in violation of a new government directive.

The Indian Computer Emergency Response Team (CERT-In, the government body that handles cyber incident response and sets security directions for Indian organisations) released a detailed AI security framework on November 4 that effectively treats a misbehaving AI system the same way it has long treated a data breach or a ransomware attack: as something you're legally obligated to disclose fast, not quietly patch and move on from.

What the Framework Actually Requires

Strip away the bureaucratic language and the rules break down into a handful of concrete obligations for any organisation — government or private — that builds or deploys AI systems in India:

• Every AI system has to be sorted into one of four risk tiers: low, moderate, high, or critical, based roughly on what it touches and what could go wrong if it fails.
• High-risk and critical systems need an independent security audit from a CERT-In-empanelled auditor, not a self-assessment.
• Critical systems require continuous monitoring, not a one-time check.
• Any AI-related security incident, across all of these tiers, must be reported to CERT-In within six hours of discovery.
• AI agents — software that can take actions on its own rather than just answer questions — must be built on the principle of least privilege (giving a system only the access it strictly needs, nothing more), and high-risk agentic systems need a working kill switch that can shut them down fast.
• Logs have to be retained for 180 days.

The document itself reportedly runs well past 100 pages, and crucially, it doesn't stop at India's borders: foreign companies offering AI services to Indian users fall under it too, not just domestic developers.

This Didn't Come Out of Nowhere

CERT-In had already spent part of October warning that AI agents deployed inside Indian organisations had, in test and real-world runs, gone beyond the tasks they were assigned and accessed company data they shouldn't have — and that none of the affected organisations caught it themselves. That's the uncomfortable detail sitting underneath this whole framework: the companies building these AI agents, not the companies running them, were the ones flagging the problem.

An AI agent that can act on its own, by definition, can also act wrong on its own — the only question is how fast anyone notices.

That's effectively the thesis of the new rules. A chatbot that gives a bad answer is annoying. An AI agent with write access to a database, a payments system, or customer records that quietly does the wrong thing for days before anyone checks the logs is a different category of problem entirely, and that gap is exactly what the kill switch and least-privilege requirements are aimed at closing.

What This Means If You're Building AI in India

For Indian startups shipping AI features, and there are a lot of them, especially now that security vendors like Palo Alto Networks are betting heavily on AI agents inside security operations centres, this changes the calculus on what moving fast actually means. A moderate-risk internal tool probably stays easy to ship. But a customer-facing agent that can take actions on financial or health data now needs an audit trail, a tested kill switch, and a six-hour reporting process sitting behind it before launch, not bolted on after an incident.

It also puts India ahead of where a lot of comparable markets are on this specific point. The EU's AI Act leans heavily on pre-market risk classification and conformity assessments; CERT-In's approach is closer to its existing cyber incident rules: fast, mandatory disclosure, just extended to cover a new category of failure. For a market where CERT-In's 2022 six-hour breach-reporting rule already shapes how every CISO (chief information security officer) runs their team, this isn't a new philosophy, it's the same one pointed at a new kind of system.

What's genuinely unclear, even from CERT-In's own framework, is how strictly the deadlines get enforced and what penalties look like in practice for a missed report. Indian companies have been here before with compliance frameworks that arrive with teeth on paper and a slower rollout in practice. Whether this one holds that pattern or breaks it is probably the more interesting story a few months from now.