NIST adds agentic AI attacks to adversarial machine-learning taxonomy
The taxonomy will matter when procurement reviews ask whether an agent's tool access survived actual adversarial assessment.
TL;DR
The National Institute of Standards and Technology (NIST) published a Nov. 9 draft update to its adversarial machine learning taxonomy, adding attacks and mitigations for agentic artificial intelligence systems with comments due Dec. 18. Certified Third-Party Assessment Organizations (C3PAOs), assessors, executives, and independent software vendors get an early NIST vocabulary for systems whose attack surface expands through tool access, memory, outside data, or other agents. The useful shift is NIST drawing a boundary around agents as a distinct attack surface.
NIST's Nov. 9 draft treats agentic systems as a separate security case, which is the point. The draft Special Publication 2400-102e2026, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, focuses on attacks on Agentic Super Intelligence (SI) systems and corresponding mitigations, with comments open through Dec. 18, 2026, according to NIST: https://www.nist.gov/news-events/news/2026/11/adversarial-machine-learning-taxonomy-and-terminology-draft-nist-sp-2400. That matters because the agent is the part of the system allowed to act. It can call tools and carry memory across sessions. It can pull context from outside sources. It can interact with other agents and accumulate permissions while looking like just another AI feature in a procurement slide deck.
NIST's own framing is careful. The document does the plumbing work of common language rather than product certification. It updates the prior adversarial machine learning work with an Agentic SI Attacks Taxonomy Index and asks commenters to identify newly relevant attacks and mitigations, plus terminology that needs standardization. That is dry standards work. It is also where procurement language starts, because solicitation clauses and assessment questions tend to inherit NIST words before anyone agrees on how much those words should require.
For Monday morning, the audience is broader than the AI lab. A Certified Third-Party Assessment Organization (C3PAO) looking at a defense supplier's AI-enabled workflow will need a vocabulary for prompt injection that arrives through retrieved content or tool results. An executive approving an internal agent needs to know whether long-horizon tasks change the risk model. An independent software vendor trying to sell into government will eventually be asked how its mitigations map to NIST terms. If that answer doesn't match the taxonomy, the review slows down while everyone translates.
The open issue is timing. The comment period closes Dec. 18, and NIST has not, on the supplied record, given a final publication date or said when the taxonomy should become normative in AI security procurement language. Even so, this is NIST moving while agentic systems are being named inside pilots and buying processes. For compliance teams, that is the narrow window when vocabulary can shape design before it becomes audit evidence for choices already made.
Published ·Deep Fathom