Back to MITRE ATLAS
MITRE ATLAS · AML.TA0010
Exfiltration
“The adversary is trying to steal ML artifacts or other information about the ML system: training data, weights, prompts.”
ATLAS tactic AML.TA0010
modeldeployment
Type
tactic
Parent
—
Sub-items
3
Sub-items
- AML.T0044Extract ML modelAdversaries query a deployed model to reconstruct an approximate copy or recover its parameters.
- AML.T0024Membership inferenceAdversaries determine whether a specific data record was part of the training set, with privacy implications.
- AML.T0057LLM data leakageAdversaries elicit memorized training data from a model via crafted prompts.
Cross-framework relationships
Lattice uses an explicit, four-type vocabulary and refuses ‘equivalent.’ Each relationship below carries a rationale and an editorial reviewer credit.
No cross-framework relationships have been recorded for this item. Either editorial review is pending, or no other framework speaks here.
Version: Pinned 2026.Q1Last reviewed: 2026-05-02