Back to MITRE ATLAS View canonical source
MITRE ATLAS · AML.T0020
Poison training data
“Adversaries inject crafted samples into a training dataset to influence model behavior.”
ATLAS technique AML.T0020
datamodel
Cross-framework relationships
Lattice uses an explicit, four-type vocabulary and refuses ‘equivalent.’ Each relationship below carries a rationale and an editorial reviewer credit.
informs
1- Validate training dataMITRE ATLAS · AML.M0007
Data poisoning (technique) is mitigated in part by training-data validation (mitigation). Within ATLAS, this is the mapped relationship.
Reviewer: Editorial: K. Park · Reviewed 2026-05-02
Version: Pinned 2026.Q1Last reviewed: 2026-05-02