Cross-Modal Misdirection
多模態交叉誤導
Verification & Validation
Risk Description
When an attacker supplies inputs in which the image and the textual description contradict each other; due to unmitigated control gaps, the system trusts the incorrect modality and executes subsequent operations accordingly, triggering external stakeholder impacts and causing because each individual modality appears normal, protective mechanisms raise no alert.
Framework Mappings
MITRE ATLASAML.T0043
NIST AI 100-2Evasion
OWASP Top 10 for LLMLLM01
ISO/IEC 5338驗證與確效
MIT AI Risk RepositoryDomain 2
Risk Treatment & Implementation Guidance
Check cross-modal consistency, pausing automation when image and text conflict; Validate each modality independently before fusion so an erroneous modality cannot dominate; Require human confirmation for high-impact follow-on actions, never auto-executing on conflicting inputs