More organisations now use AI agents: AI that takes actions, such as reading email, searching systems or running tasks. The US National Institute of Standards and Technology (NIST) has published a draft update of its guide to how AI systems are attacked. This edition focuses on agents.
NIST guidance is voluntary for most organisations, and this is a draft. There is nothing to comply with yet. Anyone can comment until 18 December 2026.
What’s changing
NIST published the initial public draft of SP 2400-102e2026, “Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations”, on 9 November 2026. “Adversarial machine learning” means deliberately attacking AI systems, or the data they learn from. The authors are from NIST, Northeastern University, ElevenLabs and the UK AI Security Institute.
The draft builds on NIST’s earlier AI 100-2 series, most recently AI 100-2 E2025. It sets out:
- A shared vocabulary. It sorts attacks by the kind of AI system, when in its life it is attacked, and what the attacker wants and knows. A glossary defines the key terms.
- Attacks and what reduces them. Its keywords include data poisoning, evasion, privacy breaches and abuse. Poisoning means corrupting the data a model learns from. Evasion means crafting input that makes a model get things wrong.
- A focus on agents. NIST says this edition concentrates on attacks on AI agents, and the ways to reduce them, in Section 4. It adds a new index of attacks on agents alongside the updated index of attacks on AI systems.
- Open problems. It describes challenges across the life of an AI system that still lack good answers.
NIST asks reviewers three questions. What are the latest attacks, especially on AI agents? What mitigations work against them? Is there new terminology that needs standardising?
The draft is meant to inform other standards and practice guides for managing AI security. Comments go by email to the address on NIST’s publication page by 18 December 2026.
What it means for you
The draft is a reference, not a checklist. It gives security teams and AI developers a common way to name the risks. Some ways to use it:
- Read Section 4 if you use AI agents. Check whether the attacks it describes could reach the agents you run.
- Use its terms in your risk register. Shared names make risks easier to compare and explain.
- Comment if you can. If you build or run AI systems, NIST is asking for real attacks and real mitigations.
These RiskWitness Blueprint requirements cover the same ground:
- RW-AIM-10 AI built in only after a risk assessment (Established). Assess the risks before building AI into a product, service or business process. The taxonomy can prompt that assessment.
- RW-AIM-11 AI agents reach only what they need (Established). Give each AI agent access only to the systems and data its task requires. That limits what an attack on the agent can do.
- RW-AIM-12 People approve agents’ serious actions (Established). Require a person to approve agent actions that would be costly or hard to undo, such as payments or deleting data.
- RW-AIM-14 AI tested against manipulated input (Advanced). Test whether crafted input, such as instructions hidden in a document or web page, can make an AI system misbehave.
How RiskWitness helps
“Which AI systems and agents do we run?” Assets and Suppliers. Record each AI system or agent as an asset with an owner. Note where it runs and which systems and data it can reach. Record the AI providers behind them as suppliers, with assurance evidence and questionnaires.
“What could go wrong, and who owns it?” Risks. Add risks for the attacks that matter to you, such as an agent tricked by hidden instructions. Give each an owner, score it before and after controls, and choose a treatment.
“What will we change?” Controls and Actions. Add controls such as limited agent access and human approval for serious actions. Map each to the requirements above, and record tests and their results. Give each gap an action with an owner and a due date.
“Does this fit the frameworks we work to?” Frameworks. The RiskWitness Blueprint is built in, with starter controls mapped to it. Any other framework you’re licensed to use can be added as one of your own by pasting in its requirements. One control can count towards several frameworks.
“What do we tell leadership?” Reports. Produce a PDF report or slide deck.
Every step is kept in a history that shows who did what and when, with the old and new values. Nobody can edit or delete it. When leadership, an auditor, a customer or a regulator asks how you manage the risks of AI, the answer is already written down.
Sources
- Adversarial Machine Learning Taxonomy and Terminology | Draft NIST SP 2400-102e2026, NIST, 9 November 2026
- NIST SP 2400-102 E2026 (Initial Public Draft): Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST, 9 November 2026
The photo is from NIST’s website. NIST has not reviewed this article and does not endorse RiskWitness.
