Learn / AI security fundamentals
Lesson 4 of 58 minReviewed
MITRE ATLAS: describing attacks on AI systems
Read ATLAS tactics, techniques, mitigations and case studies, and use ATLAS ids next to ATT&CK in detections, intelligence and red-team reports.
The OWASP lists describe classes of weakness. MITRE ATLAS (Adversarial Threat Landscape for AI Systems) describes what adversaries do: it is a public knowledge base of tactics, techniques and procedures that target AI systems [1]. If you already use ATT&CK, the shape is familiar, and ATLAS tactics carry an external ATT&CK reference where one applies [2].
The objects
| Object | What it tells you |
|---|---|
| Tactic | The adversary's objective, the "why", such as Reconnaissance or Exfiltration |
| Technique / sub-technique | A reusable method, the "how", linked to one or more tactics; sub-techniques specialise a parent |
| Mitigation | A defensive measure linked to the techniques it addresses, categorised as Policy, Technical - AI or Technical - Cyber |
| Case study | A documented instance with evidence of how techniques were applied, typed as Incident (real world) or Exercise (demonstration) |
Two technique fields help with prioritisation. Platforms say where a technique applies: Predictive AI, Generative AI, Agentic AI or Enterprise. Maturity records the evidence behind it: Feasible, Demonstrated or Realized [2]. A Realized agentic technique deserves detection work before a Feasible predictive-AI one, unless you run that kind of model.
Techniques you will meet in this hub
- AML.T0051 LLM Prompt Injection: the ATLAS view of LLM01.
- AML.T0053 AI Agent Tool Invocation: an adversary gets an agent to call its tools on their behalf.
- AML.T0110 AI Agent Tool Poisoning and AML.T0109 AI Supply Chain Rug Pull: hostile tool definitions, at install time or after approval.
- AML.T0086 Exfiltration via AI Agent Tool Invocation: the outbound leg of the lethal trifecta.
- AML.T0034 Cost Harvesting: running up someone else's model bill, the core of LLMjacking.
Using ATLAS next to OWASP and ATT&CK
The three vocabularies answer different questions. OWASP names the weakness in your system (LLM06 Excessive Agency). ATLAS names the adversary behaviour that exploits it (AML.T0053). ATT&CK names the conventional host and cloud behaviour around it, such as credential access or log tampering. A detection for an attacker disabling model invocation logging, for example, is best tagged with both an ATLAS technique and the ATT&CK technique for impairing defences, because the SOC that triages it thinks in ATT&CK.
- Detections. Tag each rule with the ATLAS technique it observes so coverage gaps show up per tactic, not per product.
- Threat intelligence. When a report describes an AI-enabled intrusion, map the steps to ATLAS so you can compare it with earlier case studies.
- Red-team findings. Record the ATLAS technique sequence that led to impact; it doubles as the scenario for a regression test.
Keeping up
ATLAS publishes monthly content updates, and the data repository distributes machine-readable releases including STIX output [2]. Pin a release in tooling and review new techniques when you update, the same way you would for ATT&CK. For the academic side of the taxonomy, NIST AI 100-2 E2025 gives a shared terminology of attacks and mitigations across the AI life cycle, useful when a research paper and a vendor report use different names for the same thing [3].
Try it on ThreatDeck
- AML.T0034 Cost Harvesting See the detections and stories linked to the LLMjacking technique.
- Bedrock invocation logging deleted An example rule tagged with ATLAS and ATT&CK ids side by side.
- Technique library Filter by MITRE ATLAS to browse tactics and techniques with linked news.
Self-check
4 questionsPick an answer to see whether it is right and why. Only your score is saved, in this browser.
-
ATLAS types case studies as Incident (real-world occurrence) or Exercise (deliberate demonstration or evaluation). The distinction matters when you judge how likely a technique is.
-
Maturity is Feasible, Demonstrated or Realized. Platforms says where a technique applies, not whether it has been seen.
-
ATLAS captures the AI-specific context and ATT&CK the conventional defence-evasion behaviour the SOC already triages. Both help the analyst.
-
OWASP names weakness classes in the system; ATLAS names adversary tactics and techniques, which is what detections and intelligence reports need.