CERT-In sets audit checks for autonomous AI agents

AI agent audits must examine unauthorised tool use, access beyond approved permissions and unsafe task delegation under security guidelines set out by the Indian Computer Emergency Response Team. The framework addresses systems that can plan tasks, use external tools and assign work to other artificial intelligence systems.

CERT-In identifies autonomous execution, retained memory and delegation as capabilities requiring closer scrutiny. Such systems can perform several steps, access information, run code or communicate externally while completing an assignment.

Auditors are asked to test whether an agent can bypass approval arrangements, misuse credentials or exercise permissions it does not hold. They must also examine whether emergency-stop controls work effectively.

Tracing memory and delegated actions

The guidelines extend AI agent audits to retained information. Checks should establish whether memory can be manipulated, whether one user's information can reach another, and whether sensitive material is kept or reused without permission.

For systems involving several agents, each agent needs a distinct identity and defined access rights. Communications between agents must be authenticated and recorded so organisations can determine who requested an action and whether that request was authorised.

Controls must also detect and restrict repeated delegation, uncontrolled task creation and loops in which agents repeatedly pass work between themselves. Records should let auditors reconstruct requests, delegated tasks, permissions, decisions, tools used and resulting external actions.

For high-risk systems, organisations need the ability to halt a single agent or an entire workflow and withdraw its credentials. The framework also calls for an inventory covering production models, training data, software and external tools.

Where outside model providers withhold internal information, organisations must record those gaps and their effect on the audit. Auditors who use AI must verify its output independently, because generated summaries cannot substitute for evidence or professional judgement.

Also read: CERT-In outlines four risk tiers for AI security audits | Binance hack disclosure deepens Bitcoin fall below $77,000