By Silas Spire • Tuesday, November 10, 2026
NIST opened a public comment period Monday on a new draft taxonomy for adversarial machine learning, with particular attention to attacks on agentic systems. The document, Special Publication 2400-102e2026, is an initial public draft. Comments are due December 18, according to both the publication record and the agency’s November 9 announcement.
The immediate product is a language for discussing security problems. NIST says the terminology is intended to inform future standards and practice guides. That is a consequential distinction for anyone buying or deploying these systems: the release is a request for technical scrutiny, rather than an announcement that a product has passed a government security test.
The publication record identifies three specific areas where NIST wants help. It asks about the latest attacks threatening these systems, effective mitigations and terminology that needs standardization. The listed authors include researchers from NIST, Northeastern University, ElevenLabs and the United Kingdom’s AI Security Institute. The agency says this edition builds on the earlier AI 100-2 series.
The draft organizes its subject around the kind of machine-learning system involved, the point in its life cycle where an attack occurs, and the attacker’s aims, capabilities and knowledge. Its contents distinguish predictive, generative and agentic systems. Separate sections address evasion, poisoning, information extraction and other classes of attacks, while the agentic material receives a dedicated taxonomy index for more precise references.
Why should ordinary readers care about a taxonomy? Consider a vendor describing two unrelated failures as an “AI attack.” Without a shared vocabulary, customers cannot easily compare the failure, determine what the attacker controlled or ask whether a proposed defense actually addresses the same problem. A clear classification gives an interviewer a way to move from a sweeping promise to a testable question. That is my assessment of the practical value of this standards work.
NIST describes the intended audience as people responsible for designing, developing, deploying, evaluating and governing these systems. Its announcement says the report draws on a literature review, conversations with specialists and the agency’s own research. Those are inputs to an evolving draft; they do not remove the need to challenge gaps or outdated assumptions during the comment period.
The document also warns that identifying commercial equipment or software does not imply NIST endorsement. Buyers should carry that distinction into discussions with suppliers. The next meaningful milestone is the close of comments and subsequent revisions. For now, technical teams have a concrete document against which to identify missing attack categories, unclear terms and defenses that need stronger supporting evidence.
Sources
FOLLOW THE ENTIRE CREW
Daily reporting and newsroom updates from FrontPage Crew.
