NIST updates its adversarial machine learning taxonomy

NIST published NIST AI 100-2e2025, an update to its voluntary taxonomy and terminology for adversarial machine-learning attacks and mitigations.

By Hypler editorial team · Source: NIST.

Original source date: .

Hypler publication: . Hypler updated: .

Estimated reading time: 4 minutes.

Topics: NIST, adversarial machine learning, AI security, taxonomy

Generated editorial illustration. Adversarial ML attack classes illustrated as contaminated data, distorted paths and escaping fragments; not a scientific apparatus.
Generated editorial illustration. Adversarial ML attack classes illustrated as contaminated data, distorted paths and escaping fragments; not a scientific apparatus. Credit: Hypler / AI-generated editorial illustration. Hypler editorial asset; not vendor or event photography

The publication supplies vocabulary, not certification

NIST announced NIST AI 100-2e2025 on March 24, 2025 as a voluntary taxonomy and terminology update for adversarial machine-learning attacks and mitigations across predictive and generative systems. It covers evasion, poisoning, privacy, and misuse categories and was developed with input from U.S. and U.K. AI security institutes, industry, and academia. The publication can help security, engineering, and governance teams describe concrete classes of concern. It does not demonstrate a deployed defense, establish complete protection, or certify an application that cites its terms.

Map categories to reachable architecture

An actionable threat model links taxonomy terms to assets, trust boundaries, attacker capabilities, controls, telemetry, and response owners. For a tool-using assistant, model prompt injection in retrieved material, unauthorized tool invocation, sensitive-output disclosure, model extraction, and training-data integrity separately. Identify what an adversary can control: training examples, retrieved documents, request text, uploads, tool output, or access to a model endpoint. Then state the effect being sought and the component that could enable it. Avoid assigning every category to every system by default; a source-aware assistant and a training pipeline expose different paths.

A hostile document tests the action boundary

Consider a hypothetical knowledge assistant reading partner documents through an MCP server. A hostile document instructs the model to export customer records. The server exposes only a tenant-filtered read query, the policy layer denies export, and telemetry records attempted instruction and decision. The response process can update fixtures and controls. This is useful because it tests the complete path from untrusted content to external effect. A prompt filter alone might reduce one phrasing while leaving tool authority unchanged; the enforcement boundary must hold even if the model repeats the hostile instruction.

Poisoning needs lineage and change control

Keep source identity, ingestion history, and review state available for datasets that affect training or retrieval. An unexpected distribution change may be a data-quality defect rather than evidence of attack, so an investigation should distinguish intent from observed behavior. A response might quarantine a batch, restore a reviewed source revision, or repeat evaluation; record why it was taken. A model score cannot identify which upstream records changed or whether they entered under expected rules. This is why data lineage is a security control as well as a maintenance record.

Evasion and extraction require different experiments

Evasion examines how crafted input changes output; extraction concerns information or behavior obtainable through model-interface access. Define an approved environment and request budget before testing either. Use synthetic material where private data is not necessary to establish a finding. Capture conditions and observed result without inflating a bounded success into a claim that every related attack works. A failed attempt also remains a result under those conditions, not a general declaration that no vulnerability exists. This precision keeps a useful taxonomy from turning into vague risk language or unsupported claims.

Mitigations need tests at their intended boundary

Maintain fixtures by category, include benign near-misses to measure false positives, and test the whole path from untrusted input to effect. Inspect logs for sensitive content, run access-control tests independently of model behavior, and repeat them after model or prompt upgrades. An input filter may block a known fixture while harming legitimate tasks or leaving another ingestion route open. For retrieval, inspect admission, source permissions, citation handling, and tool boundaries separately. Hypler can turn categories into reproducible tests and named work items; the resulting evidence supports a decision about a defined architecture, not a certification. Residual risk should remain explicit when coverage is incomplete for a named release decision.

Hypler analysis

Hypler’s view

A shared taxonomy helps security teams identify real attack paths. It becomes useful only after terms are mapped to actual assets, trust boundaries, attacker control, telemetry, and response owners in a defined architecture.

Hypler technical analysis

Engineering notes

A shared taxonomy lets teams convert broad AI-security concern into testable attack classes across predictive and generative systems. Evasion, poisoning, privacy, and misuse risks must be connected to the actual model, data pipeline, tool path, user role, and expected failure consequence rather than logged as abstract labels.

What to inspect

Threat map
Map each relevant taxonomy class to an asset, attacker capability, entry point, trust boundary, detection signal, and accountable risk owner.
Data pipeline test
Inspect ingestion, labeling, retrieval, training, and update paths for manipulated or sensitive inputs; test failure handling and provenance retention.
Tool misuse case
Exercise adversarial instructions or untrusted content that reaches a tool-using workflow; inspect containment, authorization checks, logs, and human escalation.

Scope and limits. A voluntary terminology update supports analysis. It is not a risk assessment, a complete mitigation catalog, or evidence that an individual AI system resists the attack classes named.

Terms and definitions

MCP
Model Context Protocol, an open protocol for connecting AI applications to tools, data sources, and prompts through defined interfaces.
Prompt Injection
An attack where untrusted content tries to alter an AI system’s instructions or cause unsafe tool use.
Evaluation
A repeatable method for measuring a system against defined quality, safety, or task criteria.
AI Governance
Policies, roles, controls, and review processes that guide how AI systems are selected, used, monitored, and retired.

More on this subject

Generated illustration of evidence records passing through a permission gate; not a compliance certification.
Generated illustration of evidence records passing through a permission gate; not a compliance certification. Credit: Hypler / AI-generated editorial illustration. Hypler editorial asset; not vendor or event photography

Original source

Historical source briefing with Hypler engineering analysis. Source details reflect the original publication date; implementation examples are hypothetical. Technical analysis expanded October 24, 2026; this does not change the original source event or establish a new source verification. Editorial prose revised October 26, 2026; original event, publication and source-check dates are retained.