India’s new CERT-In AI security audit guidelines 2026 provide enterprises with a framework for assessing artificial intelligence systems throughout their lifecycle. Published on November 4, the framework introduces four risk tiers, nine audit domains, six-hour reporting for reportable AI security incidents, 180-day log retention, and specific security expectations for Generative AI, RAG and Agentic AI systems.
For businesses, the guidelines extend AI security assessments beyond conventional software security. The framework brings governance, model risk, data privacy and third-party dependencies into the same audit process. This gives organizations a structured approach to examining AI deployments, including systems capable of performing increasingly autonomous tasks.
What are the CERT-In AI security auditing guidelines 2026?
The Indian Computer Emergency Response Team, or CERT-In, published its Artificial Intelligence Security Auditing Guidelines on November 4, 2026. The framework addresses the secure, accountable and auditable deployment of artificial intelligence rather than focusing exclusively on conventional cybersecurity controls.
One feature is its risk-based approach to AI security auditing. AI systems are classified into four levels: Low, Moderate, High and Critical. This allows organizations to assess security requirements according to the potential risks associated with each system instead of applying identical audit procedures to every AI application.
| Area | CERT-In guideline |
|---|---|
| AI risk classification | Low, Moderate, High and Critical |
| Audit structure | Nine core audit domains |
| Incident reporting | Within six hours for reportable AI-related security incidents |
| Security logs | Minimum retention of 180 days |
| Advanced AI architectures | Generative AI, RAG and Agentic AI explicitly covered |
Why do the four AI risk tiers matter?
The Low-to-Critical classification establishes a risk-based approach to AI governance. A limited AI implementation and a system capable of making consequential decisions may present different security risks. The framework allows those differences to influence how each system is assessed.
For enterprise teams, an AI inventory may need to contain more information than a list of models and applications. Organizations need to understand where each system operates, what information it processes and which activities it can perform. These details help determine the scrutiny required under its risk classification.
Mint’s reporting on the framework also identified the four-tier structure alongside the six-hour incident-reporting and 180-day logging requirements. These provisions place greater emphasis on maintaining records for investigations and security assessments. They also require organizations to consider how quickly they can identify and report incidents.
What are the nine AI security audit domains?
CERT-In organizes the framework around nine core audit domains. The guidelines specifically include areas such as AI Governance, Data Privacy, Model Risk Management and Third-Party AI Supply Chain Assurance. These categories address different sources of risk throughout an AI system’s development and operation.
AI security risks can originate from several parts of a deployment. An organization may have strong infrastructure security while still facing weaknesses in model governance, data handling or an external AI supplier. The framework therefore requires enterprises to examine controls beyond conventional infrastructure protection.
The framework is also anchored in seven foundational principles, including Security-by-Design and Accountability. These principles establish expectations for considering security during system design and deployment. Accountability also requires organizations to examine responsibility for managing risks after an AI system becomes operational.
What does the six-hour AI incident reporting rule mean?
Under the guidelines, reportable AI-related security incidents must be disclosed within six hours. Meeting that requirement depends on an organization’s ability to identify a security event and escalate it through internal processes. Enterprises therefore need monitoring and incident-response arrangements that account for AI-related security activity.
The separate requirement to retain AI security logs for at least 180 days supports incident investigation and subsequent auditing. Organizations need sufficient records to examine activity after a security event. Short-lived operational logs may not provide the information required when an investigation begins.
For organizations deploying AI at scale, logging and monitoring need to be considered during system development. Incident procedures must also reflect the architecture of each deployment. These requirements form part of the security controls surrounding an operational AI system.
How do the guidelines address Agentic AI, RAG and Generative AI?
The CERT-In framework addresses several AI architectures being adopted by enterprises in 2026. The guidelines explicitly cover Generative AI, Retrieval-Augmented Generation, or RAG, and Agentic AI rather than treating every implementation as a conventional predictive model. This allows security assessments to account for differences in how these systems operate.
Agentic systems can go beyond generating an answer and participate in multi-step workflows. Their ability to perform actions introduces considerations around permissions and operational oversight. Organizations therefore need controls that account for what an AI agent can access and do, alongside monitoring, accountability and security-by-design.
RAG deployments introduce additional considerations because model behaviour is connected with retrieved information. Generative AI systems also require attention to model behaviour and data handling. By including these architectures, CERT-In recognizes the need for security assessments that reflect newer forms of enterprise AI.
Why is CERT-In testing Anthropic’s Claude Mythos?
Alongside the new framework, CERT-In has secured controlled access to Anthropic’s Claude Mythos model for formal experiments involving software vulnerability identification. The Indian Express reported that the agency is conducting limited, controlled experiments with the system. These experiments examine how an advanced AI model can identify weaknesses in software.
The development illustrates another application of AI within cybersecurity. Organizations need to assess and protect AI systems, but advanced models can also assist with security testing. Vulnerability identification is one area where AI capabilities are being examined for defensive security work.
Security teams therefore need to understand the risks introduced by AI deployments while assessing whether AI tools can strengthen existing security processes. CERT-In’s controlled experiments with Claude Mythos provide an example of this second application.
How does CERT-In’s framework fit with wider AI governance standards?
The 2026 CERT-In guidelines connect with existing AI governance standards and frameworks. The framework is aligned with MeitY’s India AI Governance Guidelines under 2025 and the ISO/IEC 42001 standard published in 2023. This alignment connects security auditing with broader approaches to AI governance and management.
For enterprises operating across several markets or already developing formal AI governance programmes, this creates an opportunity to coordinate related compliance activities. Cybersecurity assessments and AI management processes address different risks but can share information and controls. Organizations can examine how existing governance practices relate to the CERT-In framework.
Why Agentic AI security is becoming an enterprise priority
The guidelines arrive alongside changes in enterprise technology adoption. McKinsey’s Technology Trends Outlook 2026, published on September 15, assesses technology adoption on a five-point scale ranging from 1, “Frontier innovation,” to 5, “Fully scaled.” The scale compares technologies at different stages of adoption.
McKinsey also removed cloud and edge computing as a standalone trend from its 2026 report because those technologies have matured and become widely adopted. Meanwhile, newer AI-driven areas, including agentic systems, are receiving increased attention in enterprise technology discussions.
As AI systems move from isolated experiments into enterprise workflows, their security requirements become more closely connected to business operations. Audits need to examine how models are governed and monitored, along with their connections to other systems. The CERT-In guidelines address these concerns through risk classifications and audit requirements.
What should enterprises take away from the CERT-In guidelines?
At QuirkyByte, we see four areas enterprises should examine under the framework: classify AI systems by risk, build security into systems from the beginning, preserve sufficient logs and prepare for rapid incident escalation. Organizations also need to assess model risk, data privacy and third-party AI dependencies rather than examining cybersecurity in isolation.
Agentic AI, Generative AI and RAG are explicitly covered by India’s AI security-audit framework. Enterprises adopting these architectures need to consider how their security and governance processes apply to them. The CERT-In guidelines provide a structure for conducting those assessments as organizations expand their AI deployments.
Frequently asked questions
When were the CERT-In AI security auditing guidelines released?
CERT-In published the Artificial Intelligence Security Auditing Guidelines on November 4, 2026. They establish a structured framework for auditing AI security, governance and associated risks.
What are the four CERT-In AI risk categories?
The framework classifies AI systems into four tiers: Low, Moderate, High and Critical. These categories support a risk-based approach to AI security auditing.
How quickly must reportable AI security incidents be disclosed?
The guidelines require reportable AI-related security incidents to be disclosed within six hours. Organizations are also expected to retain relevant AI security logs for a minimum of 180 days.
Do the CERT-In guidelines cover Agentic AI and RAG?
Yes. The framework explicitly addresses advanced AI architectures including Generative AI, Retrieval-Augmented Generation and Agentic AI. These provisions make the guidelines relevant to newer enterprise AI deployments.
What standards are the CERT-In AI guidelines aligned with?
The framework aligns with MeitY’s India AI Governance Guidelines under 2025 and the ISO/IEC 42001 standard published in 2023. This connects AI security auditing with broader governance and management practices.
