MITRE ATLAS
A knowledge base of adversary tactics, techniques and case studies against machine-learning systems, with mitigations, modelled on ATT&CK. Openly licensed; techniques and mitigations are cited by identifier.
- Duties mapped
- 1
- Mappings
- 1
- Jurisdictions
- 1
- Mitigations used
- 1
- Evidence types
- 12
- Risk areas
- 15
- Recorded incidents
- 1,147
Duties are what the law asks; controls are what an organisation operates to meet them; evidence is how it shows it did. Incidents are the harms the AI Incident Database has recorded under the risk areas those controls address.
How it is structured
Tactics group techniques (AML.Txxxx) across the attack lifecycle; each mitigation (AML.Mxxxx) names the techniques it addresses.
It is the vocabulary a red team already uses, so a control expressed as an ATLAS mitigation can be tested by the people who attack it.
Legal duties by mitigation
A duty appears under every mitigation its mapping cites, so the totals below exceed the 1 distinct duties. References that name no single mitigation are grouped at the end rather than dropped.
Other references 1 duty
-
Providers of systemic-risk GPAI models must secure the model and its infrastructure
EU AI Act · European Union · cites AML.M0005, AML.M0001
Control access to models and limit artifact release.
Controls that cite this threat model
Each control is an original description of what an organisation operates. The reference is the mitigation it corresponds to, by number only.
| Control | Mitigation | Duties served | Evidence |
|---|---|---|---|
| AI literacy and role-based training programme Training programme | AML.M0018 User Training | 4 | AI training completion records, Role-to-curriculum training matrix, AI training curriculum and materials |
| AI system inventory and classification Process | AML.M0023 AI Bill of Materials | 11 | AI system register, Risk-tier classification sign-off, AI intake and classification procedure |
| Accuracy, robustness, fairness and security testing Technical measure | AML.M0008 Validate ML Model, AML.M0003 Model Hardening, AML.M0015 Adversarial Input Detection | 15 | Pre-release test report, Test plan and acceptance criteria, Release test sign-off |
| Adversarial and red-team testing for generative AI Technical measure | AML.M0020 Generative AI Guardrails, AML.M0022 Generative AI Model Alignment | 8 | Red-team exercise report, Red-team rules of engagement and scenario library, Adversarial findings tracker |
| Automatic event logging and record retention Technical measure | AML.M0024 AI Telemetry Logging | 8 | AI system event logs, Log schema and retention standard, Log integrity and retention check |
| Data governance and dataset documentation Process | AML.M0007 Sanitize Training Data | 8 | Dataset documentation sheet, Data quality and bias check report, Dataset approval for use |
| Frontier model safety and security framework Policy | AML.M0001 Limit Model Artifact Release, AML.M0005 Control Access to ML Models and Data at Rest | 9 | Published frontier safety framework, Dangerous-capability evaluation report, Threshold notification to an authority |
| Training-data provenance and copyright register Process | AML.M0025 Maintain AI Dataset Provenance | 4 | Training source register, Copyright and rights-reservation policy, Public summary of training content |
Informational only, not legal advice. Verify every claim against the linked official sources and consult qualified counsel before acting.
Frequently asked questions
- What is MITRE ATLAS?
- A knowledge base of adversary tactics, techniques and case studies against machine-learning systems, with mitigations, modelled on ATT&CK. Openly licensed; techniques and mitigations are cited by identifier.
- Does MITRE ATLAS make an organisation legally compliant?
- No. Adoption evidences a management practice, not compliance with any statute. A crosswalk shows where the two overlap so existing evidence can be reused; it does not transfer legal obligations.