Back to MITRE ATLAS
MITRE ATLAS · AML.T0015
Evade ML model
“Adversaries craft inputs intended to cause an ML model to produce incorrect outputs at inference time.”
ATLAS technique AML.T0015
deployment
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- Adversarial trainingMITRE ATLAS · AML.M0014
Evasion is mitigated in part by adversarial training within ATLAS's mitigation catalog.
Reviewer: Editorial: K. Park · Reviewed 2026-05-02
partially addresses
1- Safety and securityUNESCO · Safety & security
ATLAS evasion characterizes a safety-and-security threat; ATLAS is scoped to adversarial behaviors only.
Reviewer: Editorial: K. Park (Security) · Reviewed 2026-05-02
Version: Pinned 2026.Q1Last reviewed: 2026-05-02