Audits and evidence
NIST drafts attack taxonomy for agentic AI systems
NIST published a draft taxonomy of attacks and mitigations for agentic AI systems. Companies can use it to build audit baselines and questionnaires.
On November 9, 2026, the National Institute of Standards and Technology (opens in a new tab) published the initial public draft of Special Publication 2400-102e2026. The document is titled Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations. This update builds on previous frameworks by introducing a dedicated attack taxonomy index specifically for agentic systems.
For chief information security officers and compliance leads, this publication is a practical tool. Companies deploying autonomous artificial intelligence tools face a gap in their security audits. Traditional frameworks do not cover the specific ways an autonomous agent can be compromised. The new taxonomy provides the exact vocabulary needed to frame data protection impact assessments, build security questionnaires, and collect evidence of compliance.
Security audits rely on a shared understanding of risk. When a company undergoes a SOC 2 audit, the auditors look for controls that mitigate known threats. Until now, the threats facing agentic systems lacked a standardized classification. This made it difficult for organizations to prove they had the right defenses in place. A company might claim its artificial intelligence is secure, but without a standard list of threats, the auditor has no baseline to measure against.
The NIST draft changes this dynamic by defining the attack vectors. The taxonomy covers key types of machine learning methods, life cycle stages of attack, and attacker goals, objectives, capabilities, and knowledge.
Consider malicious tool invocation. An agentic system often has permission to interact with external databases or application programming interfaces. If an attacker manipulates the agent into calling a tool maliciously, the company needs audit logs that show exactly what happened. The provenance of the command becomes critical evidence. Security teams must be able to trace whether a database query originated from a legitimate user prompt or an injected payload. Without this evidence, an auditor cannot verify that the access controls are functioning correctly.
Memory poisoning presents a similar challenge for evidence collection. Agentic systems often retain context over long interactions. If an attacker poisons that memory, the system might leak sensitive data or make incorrect decisions in the future. Proving that a system is secure requires evidence that memory is isolated and validated. The NIST taxonomy gives auditors the framework to ask for this specific evidence during an assessment.
This standardization also transforms vendor management. When a company purchases an artificial intelligence tool, the procurement team sends a security questionnaire. Questionnaires based on older frameworks ask about encryption and access controls. Those controls are necessary but insufficient for agentic systems. A vendor might have perfect encryption while remaining vulnerable to indirect prompt injection.
With the new publication, procurement teams can update their questionnaires to ask about the specific mitigations listed in the taxonomy. They can demand evidence that the vendor tests for agent hijacking. This moves the vendor assessment from a generic security review to a targeted evaluation of artificial intelligence risks. It forces vendors to provide concrete proof of their security posture rather than vague assurances.
Data protection officers also benefit from this structured approach. Privacy regulations require companies to assess the risks of automated processing. A data protection impact assessment must document potential harms and the controls used to prevent them. The NIST taxonomy provides a comprehensive list of technical risks to include in these assessments. It ensures that the privacy team and the security team are speaking the same language when they document their compliance efforts.
Who it applies to
The framework applies to organizations that build, deploy, or audit agentic artificial intelligence systems. This includes developers creating autonomous tools and enterprise customers integrating those tools into their workflows. The publication is designed for security professionals, privacy officers, and system evaluators who need a common language to describe threats.
NIST publications are not binding laws. They function as voluntary frameworks. However, they frequently become the baseline for industry standards and commercial contracts. Regulators and auditors often look to NIST to define reasonable security practices. Companies that ignore these frameworks may struggle to demonstrate adequate security during an audit or a regulatory investigation.
What it requires
The document does not impose legal obligations. It establishes a common language for assessing and managing the security of artificial intelligence systems. The core of the update is the new attack taxonomy index for agentic systems.
The taxonomy categorizes the lifecycle stages of an attack. It defines the goals, objectives, capabilities, and knowledge of an attacker. The text specifically identifies current challenges in the lifecycle of agentic systems and describes corresponding methods for mitigating those attacks.
Organizations are expected to use this terminology to "inform other standards and future practice guides" for security management, according to the publication. By adopting the taxonomy, security teams can map their internal controls to the specific attack vectors detailed in the draft.
The framework also includes a glossary of key terms associated with the security of these systems. This ensures that when a company claims to have mitigated a specific threat, the auditor understands exactly what that claim means.
The dates
The National Institute of Standards and Technology published the initial public draft on November 9, 2026.
The public comment period for the draft is open through December 18, 2026. The agency is specifically seeking feedback on the latest attacks threatening agentic systems, effective mitigations, and new terminology that requires standardization.
What to check now
- Review existing vendor security questionnaires to see if they address vulnerabilities specific to agentic systems.
- Map your current artificial intelligence inventory to identify which tools function as autonomous agents.
- Compare your internal risk assessment templates against the new attack vectors listed in the draft.
- Verify that your audit logs capture sufficient detail to investigate security incidents.
- Update your data protection impact assessment procedures to incorporate the terminology from the new taxonomy.
Standardized terminology is the foundation of verifiable security. The introduction of a dedicated taxonomy for agentic systems allows organizations to measure their defenses against a recognized baseline. Companies that integrate these concepts into their audit processes will be better positioned to provide clear evidence of their security posture.
Sources
This article is general information, not legal advice. For your company's situation, talk to a qualified lawyer.
This article was written with AI and checked, sentence by sentence, against the sources linked above.