CERT-In Releases Artificial Intelligence Security Auditing Guidelines
The Artificial Intelligence security auditing guidelines offer a structured framework for conducting security and government audits of AI systems developed, procured, deployed, or operated by organizations in India, as well as AI systems processing Indian data or providing AI-enabled services to Indian users.
AI systems are used across sectors but pose risks such as data manipulation, bias, and security threats. To help Indian organizations identify and mitigate risks, CERT-In formed guidelines based on risk-based frameworks to ensure responsible AI deployment aligned with Indian and international standards.
The guidelines aim to standardize security audits for AI systems in India, focusing on governance, privacy, ethics, and resilience. The guidelines provide a structured framework that adopts a risk-based approach.
Risk tiering
AI systems are classified Low, Moderate, High or Critical. This is based on factors that include impact severity, how autonomous the system is, data sensitivity, scale, explainability, reversibility, how critical the sector is, and adversarial exposure.
The tier sets how deep, how often and how independent the audit is. It also sets the documentation needed, the level of human oversight, and the monitoring and reporting requirements.
The auditing guidelines were drafted by Sub-group 03 of CERT-In’s Working Group on “Indian Standards for Cybersecurity Auditing.” The document is to be regarded as a common, risk-based framework for auditing AI systems across their whole lifecycle.
Intended Users and Stakeholders
The guidelines are designed for government agencies, critical infrastructure operators, private organizations, AI developers and providers, internal and external auditors, and CERT-In empanelled auditors. They support organizations involved in AI development, deployment, oversight, and compliance activities.
Why it matters for BFSI and CNI CISOs
“Critical” tier systems will get the most intensive and most independent audits. The third-party domain will put pressure on procurement contracts with AI vendors. The agentic-AI requirements are also the most specific any Indian regulator has published on that topic so far.
Nine Audit Domains
The guidelines identify nine core audit domains:
- Governance and accountability
- Data lifecycle, quality and privacy
- Model risk management
- Ethics, bias and fairness
- Explainability and transparency
- Security, adversarial resilience and incident reporting
- Continuous monitoring, drift and validation
- Human oversight and decision accountability
- Third-party and AI supply-chain assurance
Core Structure and Scope of Guidelines
The framework emphasizes a four-tier risk classification system influencing audit scope and frequency. It introduces nine core audit domains such as governance, data quality, model risk, ethics, explainability, and cybersecurity. Special provisions address advanced AI architectures like Generative AI, RAG, and Multi-Agent systems.
Audit Methodology and Processes
The guidelines detail a risk-based audit lifecycle from planning to follow-up, aligned with international standards. Findings are classified into levels (Critical to Low), guiding remediation priorities. Continuous improvement is encouraged through a maturity model, with AI-assisted tools supporting audit activities.
How audits run
The process goes from engagement and scoping through risk assessment, evidence collection, technical testing, reporting, remediation and follow-up checks.
Findings are graded Critical, High, Moderate or Low/Observation. Each grade comes with its own escalation and remediation obligations.
A five-level maturity model lets organisations move from ad hoc practice to continuously optimised governance.
Auditors may use AI tools, but only with approved scope, validated tools and traceable evidence. AI output can’t replace sufficient evidence or the auditor’s own judgement.
Alignment
It maps to CERT-In’s blueprint on AI-assisted vulnerability exploitation, MeitY’s India AI Governance Guidelines, ISO/IEC 42001, 23894 and 27001, NIST AI RMF, MITRE ATLAS and OWASP ML/LLM guidance.
For reporting and continual improvement it ties into ISO 42001 Clauses 9–10.
Seven guiding principles
Security-by-design, accountability and human oversight, transparency and explainability, fairness, data security and privacy, auditability and traceability, and resilience with continuous assurance.
Advanced Security and Ethical Principles
Specific security measures address prompt injection, hallucination, model poisoning, and supply chain risks. The principles focus on security-by-design, accountability, transparency, fairness, privacy, traceability, and resilience. These principles underpin all aspects of AI governance and auditing practices.
Advanced AI
There are separate requirements for GenAI, RAG, agentic AI, multimodal AI, multi-agent AI and autonomous/cyber-physical AI.
These cover prompt injection, excessive agency, data and model poisoning, vector and embedding weaknesses, unsafe tool use, persistent-memory risks, inter-agent communication, and improper output handling.
Download the report summary/explainer here.
