# CERT-In requires auditing AI agents for privilege abuse

> India publishes the first auditable checklist for agentic systems, with tests for escalation, memory leakage between users and delegation loops.
> By José Maldonado · AI Agents · 2026-11-07

India's incident response center has published an audit framework for artificial intelligence systems that act on their own, and it did so in a format unusual for this field: a list of tests that an auditor can run, with a verifiable result. [CERT-In](https://www.cert-in.org.in/), the agency under the Ministry of Electronics and Information Technology, does not talk about principles or good practices; it talks about trying to exceed an agent's permissions and checking whether it succeeds.

The document arrives at a time when the evidence on the problem is no longer anecdotal. A survey by Gravitee of more than 900 executives and technical staff, reported by [TechTarget](https://www.techtarget.com/news/4-steps-to-govern-employee-built-ai-agents-from-day-one/873863/)found that 88% of companies had experienced confirmed or suspected security incidents involving AI agents in the past year, and that only 14% deploy them in production with full security approval.

The difference between that 88% and that 14% is what the Indian framework tries to measure, and that is why it is useful even for those who do not operate in India. An auditor can apply the same tests in Madrid or São Paulo without asking permission from any local authority.

## The controls an auditor can run

The first block of tests checks whether an agent can exceed the permissions assigned to it. The auditor must try to get it to use tools without approval, to bypass the existing approval mechanisms and to reuse credentials that do not belong to it. An agent that fails any of the three is poorly designed even if its normal behavior looks correct.

The second block measures insecure delegation of tasks and the effectiveness of stop mechanisms. It is not enough to have an emergency button in the documentation; you have to prove that it actually stops the agent and how long it takes. For high-risk systems, the framework also requires the ability to suspend an individual agent or an entire flow and revoke its access credentials.

The third block is the inventory. The organization has to maintain a list of the models, training data, software and external tools it uses in production, and document the gaps when a provider does not disclose the inner workings of its model.

## The problem of agent memory

The part that makes teams that already have agents in production most uncomfortable is memory. An agent that remembers previous conversations gains usefulness and carries a risk: what it remembered from one user can show up in another user's session, and the framework requires testing whether that happens.

The tests described in the document are concrete. You have to verify whether memory can be manipulated from outside, whether one user's information leaks to another and whether sensitive data is retained or reused without authorization. These are the three questions any team should be able to answer about its own implementation and that almost none answer today.

The reason almost no one answers them is that memory is usually implemented as a vector table shared among all clients. That architecture saves money and complicates isolation, and the Indian framework puts it in the crosshairs without naming it.

For multi-agent systems, the document requires that each agent have a unique identity and defined permissions, and that communications between them be authenticated and logged. The goal is to be able to answer who requested what and whether they were authorized to request it, something that in practice requires a log that almost no agent harness keeps by default.

## The incident that explains the rush

The Gravitee survey included a case that sums up the problem better than any framework. Over roughly 36 hours, a network of agents running on an engineer's account spent close to $500,000 spinning up infrastructure in a cloud. No one broke into the system, the permissions were the engineer's own, and there was no spending limit.

That episode shows that an agent's risk lies not only in what it can read, but in what it can spend, create and hire using the identity it inherits. An agent with a person's permissions is, for all technical purposes, that person, and it acts faster than any human can react.

Vendors are responding with product. CrowdStrike presented on November 5 a [detection of anomalous agent behavior](https://cxotoday.com/media-coverage/crowdstrike-extends-the-agentic-soc-to-detect-and-stop-the-agent-insider-threat/) with a kill switch that terminates them, and Sysdig made generally available a tool that instruments the agent harness to see which tools and MCP servers it calls at each step, according to the [Sysdig announcement](https://aithority.com/it-and-devops/cloud/sysdig-launches-sysdig-ai-defense-bringing-runtime-security-to-every-action-ai-agents-take/)Both proposals target the same point as the Indian framework.

## A questionnaire someone will have to answer

The CERT-In framework is voluntary for most organizations and mandatory for regulated entities in India, so it does not change the legal situation of a Spanish team. What it does do is give them a checklist they can use on Monday morning without waiting for their regulator to publish its own.

The real value lies in the asymmetry it reveals. Any team can test in an afternoon whether its agent escalates privileges or leaks memory between users, and very few have run that test. The case of [LangFlow and its two 9.8 flaws](https://observatorioblockchain.com/en/cybersecurity/langflow-cve-2026-93680-dos-fallos-criticos-agentes-ia/) shows that the attack surface is already in production, and that of [the agents that disabled their own monitoring](https://observatorioblockchain.com/en/ai-agents/openai-agentes-desactivaron-monitorizacion-seguridad/) shows that the problem is not solved by adding unlimited observability.

Meanwhile, the public debate is still arguing over which model reasons better, a different and far less urgent question than knowing who can revoke an agent's credentials at three in the morning. The reading from the crypto sector, where the [attack on agents for analysis purposes](https://observatorioblockchain.com/en/cybersecurity/chainalysis-ataque-agentes-ia-seguridad-cripto/) has already happened, confirms that the order of the questions matters.
