CERT-In’s new AI auditing guidelines makes cyber risk prioritisation necessary
By Rajnish Gupta, MD & Country Manager, Tenable India
CERT-In’s new guidelines urges organisations to classify AI models into Low, Moderate and High or Critical risk categories. This is essentially what exposure management does. It proactively offers Vulnerability Priority Rating, ranking risks that pose the greatest risk to an organisation by focusing on attack pathways that adversaries can leverage. CERT-In’s guidelines now make an approach like exposure management necessary to manage risks arising out of AI adoption.
Most organisations adopt AI faster than they govern it. A recent study shows 35% of organisations have no policies in place to manage AI and prevent shadow AI. Many also conduct a thorough security review, integrate AI into the business and never look at it again. CERT-In’s new AI Security Auditing Guidelines reiterate that neither is the right approach, especially as AI increasingly becomes an attack vector.
The framework emphasises that organisations should first identify and document every AI system in scope, with an owner, a purpose and a risk tier assigned to each one. It urges organisations to classify AI models into Low, Moderate, High or Critical risk categories. The classification factors include the decision-making autonomy of the model or autonomous agent, the severity of potential security incidents, the sensitivity, scale and exposure of impacted data, reversibility and data recoverability. The heaviest scrutiny falls on the systems that can cause the most harm.
However, shadow AI makes this classification an uphill task. In India, shadow AI added an average of US$189,000 (INR 1.79 crore) to the cost of a data breach. Employees tend to adopt non-sanctioned tools to improve their productivity and feed critical data into these systems without thinking through the consequences. These models never make it into a register or get audited.
The guidelines nudge organisations to operate under the principle of least privilege. AI agents should not automatically inherit permissions from users, tools or other agents. Hence, organisations should treat agents the way a sensible organisation treats a new hire.
Tenable’s Cloud and AI Security Risk Report 2026 makes it clear why these actions are important. Non-human identities also hold critical excessive permissions more often than human identities do, at 52% against 37%. CERT-In’s guidelines nudge organisations to address the risks arising from these excessive permissions.
This is why organisations need visibility across their AI models to identify all of them, including non-sanctioned ones, and keep them aligned with the guidelines. Without visibility, they cannot control these systems, leaving an open door for attackers. Therefore, visibility comes first because nobody can revoke access they don’t know exists.