India’s cybersecurity agency, the Indian Computer Emergency Response Team (CERT-In), has outlined security and auditing requirements for artificial intelligence systems capable of independently planning tasks, accessing external tools and delegating work to other AI systems.
The guidance focuses on AI agents, which can do more than respond to user prompts. These systems may execute multi-step tasks, interact with software applications, retain information through persistent memory and coordinate with other agents. While these capabilities can improve automation and productivity, they also introduce additional security and accountability risks.
According to the guidelines titled Artificial Intelligence (AI) Security Auditing Guidelines: Framework for Secure, Ethical, and Auditable Artificial Intelligence Systems, organisations should evaluate how AI agents access resources, exercise permissions and carry out delegated tasks.
Why AI Agents Need Stronger Security Controls
Unlike conventional AI tools that primarily generate responses, autonomous agents may take actions using connected applications, databases, APIs and other digital resources. If their permissions are poorly configured, they could potentially access information or perform operations beyond what a particular task requires.
The risks highlighted in the guidance include unauthorised tool use, privilege abuse and insufficient oversight of automated actions. Persistent memory and task delegation can further complicate security because an agent may retain information across interactions or pass work to another system.
These concerns make it important for organisations to assess not only the accuracy of an AI system but also the boundaries within which it is permitted to operate.
Permission Checks and Task Delegation Under Scrutiny
The guidelines call for testing the permissions assigned to AI agents and checking whether they can access only the tools and resources necessary for their designated responsibilities.
Task delegation also requires careful assessment. When one agent assigns work to another, organisations should verify that the delegated activity remains within authorised limits. Security reviews should examine whether permissions are properly controlled across the entire chain of agents.
This approach can help identify situations in which an agent receives excessive access or carries out an action that was not authorised by the original task.
Emergency-Stop Mechanisms and Action Records
Another important area is the ability to interrupt an AI agent when its behaviour becomes unexpected or potentially unsafe. The guidelines call for emergency-stop mechanisms that allow organisations to halt autonomous activity when necessary.
Maintaining reliable records of agent activity is equally important. Organisations should be able to examine what an agent did, which tools it accessed and how tasks were delegated. Such records can support audits, incident investigations and accountability reviews.
For businesses deploying AI-based workflows, these controls should form part of a wider governance framework. Shunyatax's financial investigation services may also be relevant to organisations reviewing financial irregularities and evidence trails, although AI security audits require specialised technical expertise.
Multi-Agent Systems Must Maintain Authenticated Communication
The guidance also addresses systems in which multiple AI agents communicate and collaborate. Such environments need authenticated and logged communications so that organisations can establish which agent sent a message, what information was exchanged and what actions followed.
Keeping records of these interactions can improve traceability and help identify problems when several agents contribute to a single workflow. Without appropriate authentication and logging, it may become difficult to determine responsibility for an unauthorised or unexpected action.
What the Guidelines Mean for Businesses
As organisations increasingly consider AI agents for customer support, financial operations, software development and internal processes, security testing is becoming an important part of deployment planning.
Businesses should review agent permissions, monitor tool usage, assess delegation controls and ensure that emergency intervention is possible. They should also retain sufficient records to reconstruct important decisions and actions.
The broader message is that autonomous AI systems need security safeguards proportionate to the level of access and independence they are given. Testing should therefore cover not only what an agent is designed to do, but also whether it can exceed its authority or behave unexpectedly.