India’s CERT-In issues AI security auditing guidelines with four risk tiers
The guidelines, released on 4 November, set nine audit domains and tie audit depth and independence to a Low, Moderate, High or Critical tier, with annual independent audits by CERT-In-empanelled auditors for Moderate-risk systems.
Summary
MediaNama reports that India's Computer Emergency Response Team, CERT-In, published its AI Security Auditing Guidelines on November 4. The document sets out a risk-based approach to auditing AI systems that organisations in India build, buy, deploy or run. It also reaches systems that handle Indian data or supply AI-enabled services to users in India.
The guidelines cover governance, security, privacy, fairness, openness, resilience and the ability to be audited, and they define nine core audit domains spanning an AI system's life from design and procurement to decommissioning. The domains include governance and accountability, data quality and privacy, model risk, ethics and bias, explainability, adversarial resilience and incident reporting, continuous monitoring, human oversight, and third-party supply-chain assurance. Extra security requirements apply to generative AI, large language models, retrieval-augmented generation, agentic, multimodal, multi-agent and autonomous systems, covering matters such as prompt injection, data poisoning, hallucinations, deepfakes and excessive agency.
Systems are sorted into four tiers: Low, Moderate, High and Critical. The tier, based on factors such as impact severity, autonomy, data sensitivity and sector criticality, sets how deep, how frequent and how independent audits must be. Moderate systems are to be audited independently each year by auditors empanelled with CERT-In, High systems periodic audits that may occur every six months, and Critical systems the strictest scrutiny with continuous monitoring. Three human oversight models are described, and people affected by AI-assisted decisions should be able to seek review.
Findings are graded Critical, High, Moderate or Low. Moderate findings call for a plan within 30 business days, carried out inside 60 days, and Low findings must be fixed within three months. Reportable AI security incidents must be notified to CERT-In within six hours, with logs kept for 180 days.