Key Takeaways

  • CERT-In’s new guidance classifies AI systems across four risk tiers, with scrutiny increasing for high-impact deployments.
  • Audits cover nine domains, including governance, privacy, resilience, human oversight, and supply-chain assurance.
  • Agentic AI receives dedicated controls for tool access, persistent memory, autonomous actions, and emergency shutdowns.

India’s cybersecurity agency is putting more structure around how organizations assess the risks created by artificial intelligence. The CERT-In AI Security Auditing Guidelines, issued on November 4, 2026, introduce a risk-based framework for AI systems developed, procured, deployed, or operated in India.

The scope also extends to systems that process Indian data or serve users in the country. That gives the guidance relevance well beyond domestic model developers. Multinational cloud providers, software vendors, financial institutions, outsourcing companies, and enterprises embedding third-party models into customer-facing products could all encounter its expectations.

At the center of the framework are four risk tiers: low, moderate, high, and critical. Audit depth, frequency, independence, and human-oversight expectations rise with each tier. A limited internal productivity assistant, for example, would generally present a different risk profile from an autonomous system that can approve transactions, access sensitive records, or make decisions affecting essential services.

That tiered approach could help organizations avoid treating every AI application as equally dangerous. It also places pressure on governance teams to document why a system belongs in a particular category. Who owns that decision when a general-purpose model is reused across several business functions? In practice, classification is likely to involve security, legal, privacy, procurement, and business leaders rather than a single technical team.

The guidelines organize audits around nine domains: governance, data quality, security, privacy, fairness, transparency, resilience, human oversight, and third-party supply-chain assurance. As MediaNama reported, the framework combines tiered assessment with expectations around oversight and incident reporting.

Conventional software testing does not fully capture the behavior of generative or agentic systems. An application can pass standard vulnerability scans while remaining susceptible to manipulated prompts, poisoned retrieval data, or unsafe actions initiated through connected tools.

CERT-In therefore identifies advanced-AI tests for prompt injection, sensitive-information disclosure, data and model poisoning, hallucination, unsafe tool use, model extraction, insecure APIs, and unauthorized model modification. These checks move AI auditing closer to adversarial security testing, rather than limiting it to model accuracy, documentation, or bias reviews.

Agentic AI receives particular attention. The guidance covers least-privilege authorization for tools, poisoning of persistent memory and vector databases, authentication between agents, limits on autonomous actions, and tamper-resistant emergency stops. Those controls reflect a basic shift in exposure. A chatbot that produces a flawed answer creates one kind of problem. An agent that can send messages, change records, invoke code, or initiate payments creates another.

The framework emphasizes supply-chain assurance. Many enterprises do not train foundational models themselves. They assemble services from model providers, open-source components, retrieval systems, plugins, APIs, and orchestration layers. Auditors will consequently need evidence about model provenance, dependency controls, vendor access, updates, and monitoring across that chain.

That could create opportunities for AI security vendors such as Protect AI, HiddenLayer, and Robust Intelligence, which operate in areas including model supply-chain security, adversarial testing, and runtime protection. It may also push established cybersecurity providers to connect AI assessment with application security, identity governance, API security, and security operations.

The direction broadly resembles the risk-management logic found in NIST AI RMF 1.0 and the application-security concerns catalogued by the OWASP Top 10 for LLM Applications. Yet CERT-In is translating those ideas into an India-specific auditing model, a distinction likely to matter for procurement reviews and internal assurance programs. Industrial Cyber has likewise highlighted the guidelines’ relevance to organizations adopting AI within security-sensitive and operational environments.

Skills remain part of the equation. CERT-In’s CSPAI program, launched in September 2024, trains cybersecurity professionals to address AI-related threats and support trustworthy deployment. Audit frameworks have limited value if assessors lack experience testing probabilistic systems, model pipelines, vector databases, and autonomous agents.

For business leaders, the immediate task is practical: build an inventory of AI systems, map data and tool access, identify accountable owners, and preserve evidence for audits. The guidelines may not settle every technical question, but they provide organizations with a clearer basis for deciding which AI deployments deserve deeper scrutiny before their autonomy and business impact expand.