NIST AI 100-2e2025, "Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations," is a NIST Trustworthy and Responsible AI report published 24 March 2025 that defines a standard vocabulary for describing attacks on machine learning systems and their defenses. NIST's own abstract states it provides "a taxonomy of concepts and defines terminology in the field of adversarial machine learning (AML)," organized by ML method type, lifecycle stage, and attacker goals, capabilities and knowledge, so that practitioners across industry, academia and government can describe the same attack the same way.
The facts in this guide come from the NIST AI 100-2e2025 PDF itself (section numbers are given where they matter), NIST's own publication page and the CSRC record with its errata. Independent summaries such as the OECD.AI policy tracker, a Burges Salmon legal analysis and the CADE project's review describe the same taxonomy. The report does not state a total count of attack sub-categories, so this guide does not give one.
Key takeaways
- The facts in this guide come from the NIST AI 100-2e2025 PDF itself (section numbers are given where they matter), NIST's own publication page and the CSRC record with its errata.
- Before this report, researchers, vendors and regulators describing an attack on a machine learning system often used inconsistent or overlapping terms for the same technique.
- The report splits AI systems into Predictive AI (PredAI), which per its Introduction still dominates "the majority of industrial applications", and Generative AI (GenAI).
- Section 3 covers GenAI and adds a misuse objective (NISTAML.04) to availability, integrity and privacy.
What problem is this taxonomy trying to solve?
Before this report, researchers, vendors and regulators describing an attack on a machine learning system often used inconsistent or overlapping terms for the same technique. The earlier NIST AI 100-2e2023 (finalized 4 January 2024, after an 8 March 2023 initial public draft) first set out to fix that, and e2025 extends it, especially for generative AI. NIST's stated goal, per its abstract, is to "inform other standards and future practice guides for assessing and managing the security of AI systems by establishing a common language for the rapidly developing AML landscape." The report is explicit that it "does not establish any legal standard or any other legal requirement," and that it "does not make recommendations on risk tolerance." A shared vocabulary matters for the same reason it does in MITRE ATLAS or MITRE ATT&CK: a taxonomy is only useful once enough people use the same terms for the same things.
Along which dimensions does NIST classify an attack?
The Executive Summary lists five: "(i) the AI system type, (ii) the stage of the ML life cycle process in which the attack is mounted, (iii) the attacker's goals and objectives in terms of the system properties they seek to violate, (iv) the attacker's capabilities and access, and (v) the attacker's knowledge of the learning process and beyond." Knowledge runs from white-box (full knowledge of training data, architecture and hyperparameters) through gray-box to black-box (query access only), per §2.1.4.
Capabilities are where device teams can map the taxonomy onto their own attack surface (§2.1.3 for PredAI, §3.1.3 for GenAI):
| Capability | Used for | Device example |
|---|---|---|
| Training data control | Data poisoning, including backdoors | Field data uploaded for retraining |
| Model control | Model poisoning, Trojans, fine-tuning circumvention | A replaced model file in an unsigned update |
| Testing data control | Evasion, triggering a backdoor | A crafted image in front of a camera |
| Label limit | Distinguishes clean-label from regular poisoning | Labels set by your own pipeline |
| Source code control | Backdoors via modified ML code or third-party libraries | A compromised inference library |
| Query access | Black-box evasion, energy-latency, privacy attacks, prompt injection | A local API that returns confidences |
| Resource control (GenAI) | Indirect prompt injection | A document or web page an assistant reads |
How does the taxonomy classify attacks on predictive AI?
The report splits AI systems into Predictive AI (PredAI), which per its Introduction still dominates "the majority of industrial applications", and Generative AI (GenAI). For PredAI it organizes attacks by the security property violated, and the Taxonomy Index gives each class a stable NISTAML identifier you can cite in a threat model:
| Objective | ID | Example attacks (ID) |
|---|---|---|
| Availability breakdown | NISTAML.01 | Energy-latency (.014), data poisoning (.013), model poisoning (.011) |
| Integrity violation | NISTAML.02 | Evasion (.022), black-box evasion (.025), backdoor poisoning (.023), targeted poisoning (.024) |
| Privacy compromise | NISTAML.03 | Model extraction (.031), reconstruction (.032), membership inference (.033), property inference (.034) |
| Supply chain | NISTAML.05 | Model poisoning (.051) |
In plain terms: evasion crafts an input at inference time (an adversarial example) that a deployed model gets wrong; poisoning corrupts training data or the training process, sometimes leaving a hidden backdoor; privacy attacks extract information about the training data or the model. §2.1.2 notes that backdoor attacks "are the only attacks in the literature that require both training and testing data control," and that energy-latency attacks need only black-box access.
What did NIST add for generative AI and AI agents?
Section 3 covers GenAI and adds a misuse objective (NISTAML.04) to availability, integrity and privacy. Its main subsections are:
- Supply-chain attacks (§3.2): data poisoning of scraped training sets and model poisoning through maliciously designed pre-trained models. The report notes backdoors that "can persist even after downstream users fine-tune the model."
- Direct prompting attacks (§3.3): jailbreaks and information extraction, including prompt extraction (NISTAML.035).
- Indirect prompt injection (§3.4): malicious instructions in resources the model ingests at runtime, such as documents or web pages, which needs only "resource control".
- Security of agents (§3.5): an agent "iteratively prompts a model, process[es] its outputs … and provides the results back." Because agents act through tools, prompt injection can let attackers "hijack agents to execute arbitrary code or exfiltrate data." NIST calls agent-specific security research "still in its early stages."
- Benchmarks (§3.6): JailbreakBench, HarmBench, AgentDojo, and open tools such as Garak and PyRIT.
Compared with the e2023 edition, §3.5 and §3.6 are new sections. e2023 described agents only as one component of an LLM application, and the NISTAML index first appears in e2025. This is the same attack surface covered tactic-by-tactic in MITRE ATLAS and control-by-control in the OWASP Top 10 for LLM Applications; NIST supplies the vocabulary both assume.
Which mitigations does the report describe, and how far do they go?
NIST pairs each attack class with mitigations, and is blunt about their limits:
- Evasion (§2.2.5): "many of the proposed mitigations against adversarial examples have been ineffective against stronger attacks." Three classes have held up: adversarial training, randomized smoothing and formal verification. All "exhibit inherent trade-offs between robustness and accuracy" and add training cost.
- Poisoning (§2.3): training data sanitization, trigger reconstruction, model inspection, and for targeted poisoning, access controls plus "dataset provenance and integrity attestation." Targeted poisoning is "notoriously challenging to defend against."
- Privacy (§2.4.5): differential privacy bounds membership inference and reconstruction, but "does not provide guarantees against model extraction attacks."
- GenAI supply chain (§3.2.3): publish and verify cryptographic hashes for downloaded data, and understand "models as untrusted system components."
- Indirect prompt injection (§3.4.4): because "current mitigations do not offer full protection," design systems assuming injection is possible. Examples are multiple LLMs with different permissions, or untrusted data only through well-defined interfaces.
Why does this matter for devices that embed AI on-device?
Two parts of this taxonomy map directly onto device and firmware work. First, products that run an on-device predictive model sit inside the PredAI half: an anomaly-detection model in an industrial sensor, an object detector in a camera, a predictive-maintenance classifier in a gateway. An attacker with physical or network access may attempt evasion. If the device supports on-device learning or periodic retraining, poisoning becomes possible too. Once firmware is extracted, the model file itself is available, and the attacker moves from black-box to white-box knowledge. The AI-BOM for embedded AI devices guide explains how to inventory the model first: you cannot threat-model a model you have not recorded.
Second, §4.2.4 on quantized models speaks directly to edge hardware. Quantization is "a technique for efficiently deploying models to edge platforms, such as smart phones and IoT devices," typically int8 or 4-bit instead of float32. NIST warns that "quantized models inherit the vulnerabilities of the original models and introduce additional weaknesses," with error amplification hurting adversarial robustness. It also cites results showing quantization "can be exploited to produce a harmful quantized LLM, even though the full-precision counterpart appears benign." Robustness numbers measured on the float32 model do not carry over to the int8 build you ship.
Third, devices that embed a generative assistant, such as a voice assistant or an agent reading logs and sensor data, sit inside the GenAI half. A malicious instruction hidden in a scanned document or a cloud feed is exactly the indirect-injection case, and tool access turns it into the agent case. The NIST AI Risk Management Framework Map function and the NIST AI 600-1 generative AI profile ask teams to identify which of these risks apply before measuring them.
What does a taxonomy-based threat model look like for a device?
Worked example, synthetic: a fictional smart doorbell with an on-device int8 person detector and a cloud-backed voice assistant that can unlock a paired smart lock. It does not describe a real product.
| Asset | Threat (NIST ID) | Capability / knowledge | Mitigation direction |
|---|---|---|---|
| Person detector | Evasion (NISTAML.022) with a printed patch | Testing data control; white-box after firmware extraction | Robustness tests on the int8 build; do not treat detection as an access-control decision |
| Person detector | Model extraction (NISTAML.031) | Model file read from flash | Encrypt the model at rest; assume it is public |
| Model update | Model poisoning (NISTAML.051) | Model control via update channel | Sign model files like firmware and verify on device |
| Voice assistant | Indirect prompt injection (NISTAML.015) | Resource control: a calendar entry it reads | Assistant cannot call "unlock" from untrusted content; confirm on the phone |
| Voice assistant | Prompt extraction (NISTAML.035) | Query access | Keep secrets out of the system prompt |
What is a workable starting point?
Identify which half of the taxonomy appliesAn on-device classifier needs PredAI's evasion, poisoning and privacy classes; an embedded assistant or agent needs GenAI's supply-chain, prompt-injection, misuse and agent sections. Many products need both. Cite NISTAML IDs in the threat modelso findings stay comparable across teams and with ATLAS mappings. Treat direct and indirect prompt injection as separate risksIndirect injection through a sensor feed, scanned document or network response is a data-provenance problem, not just input validation. Record the model and assistant in your AI-BOMbefore threat-modelling them; see AI-BOM for embedded AI devices. Scope tests by attacker knowledgeIf firmware can be extracted, test white-box against the quantized model you actually ship. Pin the edition and check the errataNIST issued a corrected PDF on 1 April 2025. The errata of 3 June 2025 replaces the GenAI availability entries on index page x with "Manipulate Resources (ID: NISTAML.016)" and "Time-Consuming Tasks (ID: NISTAML.017)". Cite the DOI and date.
Which mistakes do teams make when they apply it?
- Treating it as a compliance checklist. It sets no controls and no risk tolerance; it is vocabulary for your own risk decisions.
- Testing the float32 model only. §4.2.4 says quantization adds weaknesses.
- Relying on a detector as the defence. NIST records detection methods bypassed in published evaluations, and LLM-based guardrails "are also vulnerable to attacks."
- Ignoring the conventional software around the model. §3.2 starts from "Since AI is software, it inherits many of the vulnerabilities of the traditional software supply chain."
- Copying the index without the errata. The GenAI availability IDs on page x changed.
Where does SoftWSP fit?
SoftWSP does not test models, run adversarial attacks or produce an AML threat model. It covers the conventional software around the model: its firmware pipeline turns an image's software inventory into a CycloneDX 1.6 SBOM (firmware SBOM) and matches the components against NVD, CISA KEV, FIRST EPSS and CISA ICS advisories. An AI-drafted, cited report then goes to an engineer for review. The model-level threats above stay with your ML and security engineers. See Device & Firmware Security for firmware vulnerability work on devices that also carry an on-device model or assistant.
This guide summarizes NIST AI 100-2e2025, its CSRC record and errata, and the prior NIST AI 100-2e2023. It is not a substitute for the current PDF and does not constitute security or compliance advice.
Put this guide to work
Frequently asked questions
What is NIST AI 100-2e2025?
NIST AI 100-2e2025, "Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations," is a NIST Trustworthy and Responsible AI report published 24 March 2025 (DOI 10.6028/NIST.AI.100-2e2025). Per its own abstract, it "provides a taxonomy of concepts and defines terminology in the field of adversarial machine learning (AML)," organized "in a conceptual hierarchy that includes key types of ML methods, life cycle stages of attack, and attacker goals, objectives, capabilities, and knowledge," alongside mitigation methods and a glossary. NIST states the purpose is to "establish a common language for the rapidly developing AML landscape" so that researchers, vendors and auditors can describe the same attacks and defenses consistently. It is a reference taxonomy, not a compliance standard or regulation — it sets no mandatory controls.
How is NIST AI 100-2e2025 different from the earlier NIST AI 100-2e2023?
NIST AI 100-2e2023 was finalized on 4 January 2024 (after an initial public draft released 8 March 2023), authored by Apostol Vassilev, Alina Oprea, Alie Fordyce and Hyrum Anderson, and already covered evasion, poisoning, privacy and abuse attacks across predictive and generative AI at a foundational level. The e2025 edition, published 24 March 2025, added two new co-authors — Xander Davies of the UK AI Security Institute and Maia Hamin of the US AI Safety Institute — reflecting international input into the update. Secondary analyses that reviewed both editions describe the 2025 update as substantially deepening generative-AI coverage: expanded treatment of large-language-model and chatbot-specific attacks, retrieval-augmented-generation (RAG) systems, and AI-agent security, which gets its own section (3.5) in e2025, while the 2023 edition described agents only as one component of an LLM application. Readers who need exact wording changes between the two editions should compare the two PDFs directly.
What attack categories does the taxonomy define for predictive AI versus generative AI?
Per NIST's own publication abstract and keywords, cross-checked against independent summaries (including a UK law firm's analysis and the CADE project's review), the taxonomy first splits AI systems into two classes: Predictive AI (PredAI), described as still dominating industrial applications, and Generative AI (GenAI), whose adoption has grown rapidly. For PredAI, the taxonomy addresses evasion attacks (fooling a deployed model at inference time), poisoning attacks (corrupting training data or the training process), and privacy attacks (extracting information about training data or the model itself). For GenAI, coverage extends to supply-chain attacks, direct and indirect prompt injection, misuse/abuse violations (such as jailbreaks), and the security of AI agents built on generative models. NIST's own keyword list for the document includes "attack taxonomy," "abuse," "data poisoning," "evasion," "attack mitigation," "large language model," "chatbot," and "privacy breach." An exact, numbered count of sub-categories is not confirmed from the primary PDF text, so this guide does not state one.
What dimensions does NIST use to classify an individual attack?
Independent summaries of the report (consistent across the sources we checked) describe attacks as classified along several dimensions together, not just by attack name: the type of AI system targeted (predictive or generative), the stage of the machine learning lifecycle at which the attack occurs (from design and training through deployment and inference), the attacker's goal or the system property being violated (availability, integrity or privacy, plus misuse for GenAI), the attacker's capabilities and access level, and the attacker's knowledge of the target system (ranging from black-box query access to full white-box knowledge of model internals). NIST's own abstract describes this as "a conceptual hierarchy that includes key types of ML methods, life cycle stages of attack, and attacker goals, objectives, capabilities, and knowledge." This multi-dimensional structure is why the same named attack — prompt injection, for example — can appear differently depending on whether it is direct or indirect, and at what point in a system's lifecycle it is introduced.
Who wrote NIST AI 100-2e2025, and is it mandatory for any device or product category?
The final 24 March 2025 edition lists six authors: Apostol Vassilev (NIST), Alina Oprea (Northeastern University), Alie Fordyce and Hyrum Anderson (Cisco), Xander Davies (UK AI Security Institute), and Maia Hamin (US AI Safety Institute) — spanning NIST's own Computer Security Division, academia, industry, and two national AI safety/security institutes. NIST issued a corrected PDF on 1 April 2025 (updating author affiliations for corporate changes and improving figure alt text) and published an errata/potential-updates document dated 3 June 2025 noting at least one identified error pending correction in a future revision. Like the companion AI RMF (NIST AI 100-1), this taxonomy is not a regulation and imposes no mandatory requirements on device makers, cloud providers or any other organization — it is a vocabulary and classification system that other standards, procurement requirements or internal security programs can choose to reference.
Sources
- NIST: Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations (publication page, NIST AI 100-2e2025)
- NIST CSRC: AI 100-2 E2025, Adversarial Machine Learning (final publication page — abstract, authors, keywords, document history, errata)
- NIST CSRC: AI 100-2 E2023, Adversarial Machine Learning (final publication page — prior edition, document history including IR 8269 and 2023 draft)
- NIST AI 100-2e2025 full text (PDF, nvlpubs.nist.gov)
- NIST AI 100-2e2023 full text (PDF, nvlpubs.nist.gov — prior edition)
- NIST AI 100-2e2025 Errata (Potential Updates), last updated 3 June 2025 (PDF)
- OECD.AI Policy Observatory: Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations (NIST AI 100-2e2025)
- Burges Salmon: Adversarial machine learning — NIST terminology of attacks and mitigations
- CADE Project: NIST report unveils the tactics behind machine learning exploits
About this guide
SoftWSP Firmware Security Desk
Covers the Firmware vulnerabilities & fixes cluster: 131 guides.

In this cluster
Questions about this page?
