---
title: What does NIST AI 100-2 say about attacks on AI agents?
url: https://opencontroller.lyzr.com/blog/nist-ai-100-2-adversarial-ml-taxonomy-for-agents
published: 2026-11-18
updated: 2026-11-18
author: Lyzr
---

# What does NIST AI 100-2 say about attacks on AI agents?

**NIST AI 100-2 E2025, "Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations", was published in March 2025. Compared with the January 2024 edition it adds a section on AI agent security, expands the list of indirect prompt injection techniques, and splits supply chain attacks into data poisoning and model poisoning.** It is a shared vocabulary, not a control standard, so use it to name threats in your agent threat model and to structure red-team tests and evidence.

## What is NIST AI 100-2?

It is a NIST Trustworthy and Responsible AI report that classifies attacks on machine learning systems and the mitigations proposed for them. It covers two families:

- **Predictive AI,** such as classifiers, with attacks like evasion, poisoning and privacy attacks.
- **Generative AI,** including large language models and the applications and agents built on them.

For generative AI, attacks are organised by the attacker's objective: **availability breakdown, integrity violation, privacy compromise, and misuse enablement.** The 2025 edition renamed the earlier "abuse" category to "misuse", covering attempts to bypass safety restrictions, such as jailbreaks.

## What changed in the 2025 edition?

Reported changes include:

1. **A new section on AI agent security,** noting that agents inherit LLM vulnerabilities and add risk through their ability to take actions.
2. **More indirect prompt injection techniques,** including self-propagating injections, where content read by a model instructs it to spread the payload further.
3. **Supply chain attacks split** into data poisoning and model poisoning subsections, reflecting reliance on third-party foundation models.
4. **Mitigations grouped by stage:** pre-training, post-training, evaluation and deployment.
5. **A new subsection on benchmarks** for adversarial testing.
6. **An expanded challenges section,** covering supply chain risk, evaluating mitigations and trade-offs between trustworthiness attributes.

## Why is indirect prompt injection the agent problem?

Direct prompt injection comes from the user. Indirect prompt injection arrives inside data the agent reads: a web page, an email, a document in a retrieval index, or a tool result. Agents read untrusted data constantly and then act on it, so this is the attack class that turns a text-generation flaw into an unauthorised action.

NIST's example of a self-propagating injection is an email that instructs the model to send a message to everyone in the user's contacts. For an agent with email tools, that is a realistic attack path, not a hypothetical.

## How do you map the taxonomy to agent tests?

| Attacker objective | Agent example | Test to run | Evidence to keep |
|---|---|---|---|
| Integrity violation | Injected instruction in a retrieved document makes the agent change a record | Plant instructions in documents and tool outputs; check whether write tools are called | Test cases, results, tool-call logs |
| Privacy compromise | Agent is induced to reveal system prompt, other users' data or secrets | Extraction prompts across turns; checks for data from other sessions | Red-team findings, redaction test results |
| Availability breakdown | Input that makes the agent loop, call tools repeatedly or burn tokens | Loop and recursion tests; budget limit tests | Rate-limit and budget configuration, alert history |
| Misuse enablement | Jailbreak makes the agent perform an action outside policy | Multi-turn jailbreak suites against action policies | Eval scores per version |
| Supply chain (data/model poisoning) | Poisoned memory, tampered MCP server or skill | Write-validation tests on memory; scan of tools and skills | Inventory with versions and scan results |

Run these on every agent version, not once. Multi-turn tests matter because many attacks need several steps. See [how to red-team AI agents with multi-turn tests](/blog/how-to-red-team-ai-agents-multi-turn).

## How does it relate to MITRE ATLAS and OWASP?

The three are complementary:

- **NIST AI 100-2** gives terminology and a classification of attacks and mitigations.
- **MITRE ATLAS** gives adversary tactics and techniques in an ATT&CK-style matrix, useful for threat modelling. See [threat modelling AI agents with MITRE ATLAS](/blog/mitre-atlas-threat-model-ai-agents).
- **OWASP's Top 10 for Agentic Applications** lists the most common agent-specific risks for builders.

A practical approach is to tag each red-team test with its NIST AI 100-2 category, ATLAS technique ID and OWASP risk. That gives one evidence set that answers questions framed in any of the three.

## Do prompt-injection detectors solve it?

NIST is explicit that there is no complete defence against these attacks today; mitigations reduce risk rather than eliminate it. Detectors help, but defence in depth matters more: least-privilege tools, separation of read and write actions, confirmation for high-impact actions, and monitoring of agent behaviour. Our analysis of [whether prompt injection detectors stop indirect injection](/blog/do-prompt-injection-detectors-stop-indirect-injection) goes deeper. For the broader control model, see [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane).

## How Opencontroller does this

Opencontroller runs multi-turn and red-team evals per agent version, scores production traffic with LLM judges, and applies live groundedness and task-adherence guardrails. Rogue-agent handling monitors behavioural deviation and can alert, quarantine or kill an agent. Policy as code ships with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001, so test results can be organised as evidence for each. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**When was NIST AI 100-2 E2025 published?**
March 2025. It updates the first edition, published in January 2024.

**Does NIST AI 100-2 cover AI agents?**
Yes. The 2025 edition adds a section on AI agent security and expands its coverage of indirect prompt injection, the main way attackers hijack agents.

**Is NIST AI 100-2 a compliance standard?**
No. It is a taxonomy and terminology report. Organisations use it to structure threat models and tests, and as evidence for programmes built on the NIST AI RMF.

## Key facts

NIST AI 100-2 E2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, was published in March 2025, updating the January 2024 edition. It classifies attacks on predictive and generative AI; generative AI attacks are grouped by objective: availability breakdown, integrity violation, privacy compromise and misuse enablement ("abuse" was renamed "misuse"). New in 2025: an AI agent security section, more indirect prompt injection techniques including self-propagating injections, supply chain split into data and model poisoning, and a benchmarks subsection. It is a taxonomy, not a control standard.

## Related

- [How do you use MITRE ATLAS to threat-model AI agents?](https://opencontroller.lyzr.com/blog/mitre-atlas-threat-model-ai-agents.md)
- [How do you red-team AI agents with multi-turn attacks?](https://opencontroller.lyzr.com/blog/how-to-red-team-ai-agents-multi-turn.md)
- [Do prompt injection detectors stop indirect prompt injection?](https://opencontroller.lyzr.com/blog/do-prompt-injection-detectors-stop-indirect-injection.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
