MITRE ATLAS Explained: Mapping AI Agent Threats to Tactics and Techniques

ASA AI Team · Updated 2026-11-08 · 4 min read

MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is a public knowledge base of adversary tactics, techniques and real-world case studies targeting AI systems. It extends the logic of MITRE ATT&CK to machine learning and generative AI, and gives security teams a common language to describe, test and detect attacks on AI.

Key takeaways

ATLAS vs ATT&CK

MITRE ATT&CK describes how attackers compromise enterprise IT. ATLAS reuses the same structure and many of the same tactics — reconnaissance, initial access, execution, persistence, exfiltration, impact — and adds tactics and techniques that only exist when the target is an AI system, such as gaining access to a model, staging attacks against it or manipulating its training data. Each technique has an identifier of the form AML.Txxxx and is linked to documented case studies.

Key ATLAS techniques for LLM applications and agents

The matrix evolves regularly; always check atlas.mitre.org for the current version. The techniques below are the ones we see most often when assessing LLM applications and agents:

TechniqueIDWhat it looks like in practice
LLM Prompt InjectionAML.T0051Instructions typed by a user or hidden in documents, web pages or tool outputs override the system's intended behaviour.
LLM JailbreakAML.T0054Role-play, encoding or multi-turn tricks that make the model ignore its safety rules.
LLM Plugin CompromiseAML.T0053Abusing the tools, plugins or connectors an LLM can call to act on other systems.
LLM Data LeakageAML.T0057Extracting sensitive data the model was given in its context or training.
Poison Training DataAML.T0020Inserting malicious examples into training, fine-tuning or retrieval data.
Craft Adversarial DataAML.T0043Building inputs designed to make a model misbehave or misclassify.
Exfiltration via ML Inference APIAML.T0024Using repeated queries to extract information about the model or its data.

Applying ATLAS to an AI agent: a worked example

Take a support agent that reads customer tickets, searches a knowledge base and can issue refunds through an API.

  1. Reconnaissance — the attacker probes the chatbot to learn its tools and rules.
  2. Initial access / execution — a ticket contains hidden instructions (indirect prompt injection).
  3. Privilege escalation through tools — the agent is convinced to call the refund API outside policy.
  4. Exfiltration — the agent includes customer data in a reply or a link.
  5. Impact — financial loss and a data breach.

Each step maps to an ATLAS technique, so each step can be tested in red teaming and covered by a detection or a control: content scanning on ingestion, tool-level authorisation, output filtering and monitoring of tool calls.

Three ways to use ATLAS

Threat modelling

Walk through the tactics for each AI component and ask which techniques apply. The output is a list of credible attack paths, not a generic checklist.

Red teaming

Tag every test case and every finding with its ATLAS technique. Reports become comparable over time and across systems.

Detection coverage

Map your guardrails, logs and SOC rules to techniques to see which ones you can detect and which ones are blind spots — the same way SOC teams use ATT&CK coverage maps.

Frequently asked questions

What does ATLAS stand for?

Adversarial Threat Landscape for Artificial-Intelligence Systems. It is maintained by MITRE and freely available at atlas.mitre.org.

Is MITRE ATLAS a compliance standard?

No. ATLAS is a knowledge base of adversary behaviour. It supports threat modelling, testing and detection, and is often referenced alongside compliance frameworks such as NIST AI RMF, ISO/IEC 42001 and the EU AI Act.

How is ATLAS different from OWASP Top 10 for LLM?

OWASP Top 10 for LLM Applications ranks the most critical risk categories for developers. ATLAS describes how adversaries actually attack AI systems, technique by technique, with case studies. They are complementary.

Can ATLAS be used in a SOC?

Yes. Tagging AI-related alerts with ATLAS techniques lets a SOC measure detection coverage for AI systems and correlate AI incidents with classic ATT&CK activity.

Related guides