Theory · 7 min
MITRE ATLAS for AI attacks
Map an AI attack to ATLAS, and know when ATT&CK covers it too.
MITRE ATLAS is ATT&CK’s sibling for attacks on AI systems: the same tactic and technique structure, plus AI-specific tactics such as AI Model Access and AI Attack Staging. Technique IDs look like AML.T0051.
ATLAS covers the model layer: prompt injection (AML.T0051), jailbreaks (AML.T0054), system prompt extraction (AML.T0056), data leakage (AML.T0057), poisoning (AML.T0020) and agent tool abuse (AML.T0053 AI Agent Tool Invocation, AML.T0086 Exfiltration via AI Agent Tool Invocation).
Since v20, ATT&CK Enterprise also has AI techniques where they touch enterprise systems: T1698 Indirect Prompt Injection and T1700 AI-Enabled Exfiltration. In a SOC report, use ATT&CK for the enterprise steps and add ATLAS for the AI-specific detail.
NIST’s AI 100-2 E2025 gives a matching vocabulary for adversarial machine learning if you need to write for researchers or auditors.
T1698T1700AML.T0051.001AML.T0054AML.T0056AML.T0020AML.T0086
Your turn: match observations. Match each step to its ATLAS technique.
Do the interactive check Free, in the browser.
Sources
- MITRE ATLAS: adversarial threat landscape for AI systems MITRE
- ATLAS data (technique IDs and names) MITRE
- AI 100-2 E2025: Adversarial Machine Learning, a taxonomy and terminology NIST
- ATT&CK T1698: Indirect Prompt Injection MITRE ATT&CK v20