CERT-In Sets Risk-Based Security Audit Standards for AI Systems

The guidelines introduce four risk categories and call for independent audits, red-team testing and additional safeguards for autonomous AI agents.

TMR Staff 4 days ago

Topics

India’s cybersecurity agency has issued risk-based security auditing guidelines for artificial intelligence systems, calling for independent assessments of higher-risk applications and stricter controls on autonomous AI agents.

The Indian Computer Emergency Response Team released its 107-page AI Security Auditing Guidelines on November 4. The framework covers organizations operating AI systems in India as well as companies handling Indian data, including overseas providers serving Indian users.

CERT-In divides AI systems into four categories based on factors including autonomy, data sensitivity, the number of people potentially affected and whether any resulting harm can be reversed.

Low-risk systems may undergo internal assessments, while moderate, high and critical systems are expected to face independent audits by CERT-In-empanelled security auditing organisations. Critical systems are also expected to be continuously monitored.

The guidelines place additional scrutiny on agentic AI systems that can plan tasks, use software tools or interact with external systems with limited human intervention.

Auditors are expected to test whether such agents can exceed their permissions, misuse credentials, bypass approval mechanisms or make unauthorised use of connected tools. Multi-agent systems should maintain authenticated communications, defined permissions and logs that allow investigators to reconstruct actions.

The framework also calls for emergency controls capable of suspending individual agents or entire workflows in high-risk environments.

Organizations running production AI systems are expected to maintain an AI bill of materials covering components such as models, datasets, prompts, software and external tools. CERT-In had previously included AI bills of materials in its broader supply-chain security guidance.

The guidelines also bring existing CERT-In incident-reporting and logging requirements into AI security audits and call for annual red-team exercises for critical systems.

The move follows a series of CERT-In warnings about AI-enabled cyber risks. In October, the agency said AI agents had gained unauthorised access to Indian ICT systems in some research and evaluation exercises, while attackers were separately using open-source agentic tools for financially motivated intrusions.

The document, however, does not specify when the new audit framework becomes applicable, the legal provision under which individual requirements are mandatory, or what penalties would follow non-compliance.