- The Indian Computer Emergency Response Team (CERT-In) has set out specific security and auditing requirements for AI systems that can independently plan and carry out tasks, use external tools, and delegate work to other AI systems.
- The Artificial Intelligence (AI) Security Auditing Guidelines identify features such as autonomous task execution, persistent memory, and the ability to delegate work as factors that can increase risks.
- The framework calls for testing whether an AI agent can exceed its authorised permissions, use tools without approval, bypass approval mechanisms, or misuse credentials.
- For multi-agent systems, each agent must have a unique identity and clearly defined permissions, and communications must be authenticated and logged to establish which agent made a request and whether it was authorised.
- Organisations must have controls to detect and limit repeated delegation, uncontrolled creation of new tasks, and loops in which agents keep assigning work to one another.
- The guidelines address risks associated with an agent's memory and retained information, including testing for manipulation, user-to-user leakage, and unauthorised retention or reuse of sensitive information.
- Auditors should examine unsafe delegation of tasks and the effectiveness of emergency-stop mechanisms.
- The guidelines call for records that allow auditors to reconstruct what happened within a multi-agent system, capturing requests, delegated tasks, permissions used, decisions, software tools invoked, and external actions.
- For high-risk systems, the framework specifies the ability to suspend an individual agent or an entire workflow and revoke its access credentials.
- Organisations must maintain an inventory of components used in production AI systems, including models, training data, software, and external tools; where third-party providers do not disclose internal details, organisations must document information gaps and their impact on audits.
- Auditors using AI must independently verify its outputs, as AI-generated summaries cannot replace audit evidence or professional judgement.
The Indian Computer Emergency Response Team (CERT-In) has introduced comprehensive AI Security Auditing Guidelines aimed at enhancing the security and ethical use of artificial intelligence systems. These guidelines focus on critical aspects such as autonomous task execution, memory retention, and task delegation among AI agents.
The framework outlines the necessity for testing AI agents to ensure they do not exceed their authorized permissions or misuse tools and credentials. It emphasizes the importance of examining unsafe delegation of tasks and the effectiveness of emergency-stop mechanisms.
Moreover, the guidelines highlight the risks associated with an agent's memory, urging organizations to test for potential manipulation of memory and the unauthorized retention or reuse of sensitive information. Each AI agent must possess a unique identity and clearly defined permissions, with all communications authenticated and logged to track authorization.
For high-risk systems, the framework mandates the ability to suspend agents or revoke access credentials. Organizations are also required to maintain an inventory of all components used in production AI systems, ensuring transparency and accountability in audits. Auditors are advised to independently verify AI outputs, as AI-generated summaries cannot substitute for professional judgment or audit evidence.
“The guidelines also address risks from persistent memory, requiring tests for manipulation and user-to-user leakage, and mandate unique identities and authenticated logs for multi-agent systems. For high-risk systems, CERT-In specifies the ability to suspend agents or workflows and revoke access credentials.”










