Science & Technology7 November 20263 min read
CERT-In issues AI Security Auditing Guidelines for agentic AI systems
The Indian Computer Emergency Response Team (CERT-In) has issued the Artificial Intelligence (AI) Security Auditing Guidelines. The framework is meant to regulate, monitor and audit the emerging risks of agentic AI systems and their autonomous capabilities.
Auditing is the central tool. Systems are sorted by risk, and the level of scrutiny follows from that sorting.
Why agentic AI is treated separately
The guidelines distinguish agentic AI from the generative AI most users know. Generative systems mainly respond to prompts with text, images or code. Agentic AI refers to autonomous systems that take independent actions to achieve designated goals.
These "agents" can interact with other software, execute commands, browse the internet and manipulate files with minimal human intervention. Because they act rather than merely answer, the security question shifts from what a system says to what it does.
How the audit framework works
The framework classifies AI systems into four risk tiers: Low, Moderate, High and Critical. Placement depends on three factors: the severity of impact, the sensitivity of the data involved and the system's decision-making autonomy. The tier then dictates the depth, frequency and independence of the audit. A system acting autonomously on sensitive data therefore faces a more demanding audit than a low-stakes tool.
Audits are organised around nine audit domains that span the entire AI lifecycle. Among them are AI Governance, Data Privacy, Model Risk Management, Bias and Fairness, Explainability, Continuous Monitoring, Human Oversight and Supply-Chain Assurance.
Newer architectures get dedicated security protocols: Generative AI, Retrieval-Augmented Generation (RAG) and Agentic AI. These protocols are meant to mitigate threats early, including prompt injection and data or model poisoning, as well as misinformation.
Keeping humans in the audit loop
The guidelines allow AI-assisted audit procedures for analysing evidence, but only within limits. Confidentiality controls, tool validation and competent human review must apply. AI outputs cannot replace an auditor's independent judgment. In effect, the framework's human-oversight logic is turned on the auditors themselves.
Fit with national policy and global standards
The guidelines are described as highly interoperable. They are designed to align with MeitY's India AI Governance Guidelines 2025, issued under the IndiaAI Mission, and with ISO/IEC 42001. Published in 2023, ISO/IEC 42001 is the world's first international standard for an AI Management System (AIMS). The International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC) developed it jointly.
The framework rests on seven guiding principles described as vital to India's digital policy. These include Security-by-Design, Accountability and Human Oversight, Transparency, Fairness, Data Privacy and Auditability.
The result is a move from broad governance principles to an auditable, risk-tiered security framework, built around the autonomous, action-taking nature of agentic AI.