MITRE ATLAS · Defense Evasion
AML.T0129 · Triggers in Multimodal Inputs
Adversaries may place instructions or triggers in one part of a multimodal input to influence the model while staying unnoticed by human reviewers and by defenses that do not inspect all input modalities.
Mitigations
- Validate AI Model
- Generative AI Guardrails
In the news
No stories mapped to this technique yet.
Mappings are AI-assisted and checked against the framework definitions; corrections welcome.