S0124RS3-T10-S0124-Z · Full risk code

Multimodal Hidden Instruction Injection

多模態隱藏指令注入

Verification & Validation
Risk Description

When a user uploads an image file that appears normal but contains instruction text barely discernible to the human eye; due to unmitigated control gaps, the system reads it and executes the instructions, while the organization's text-layer filtering is never triggered, triggering compliance exposure and operational reputational costs.

Framework Mappings

OWASP Top 10 for LLMLLM01
MITRE ATLASAML.T0051
NIST AI 600-1Information Security
ISO/IEC 5338運作與監控
MIT AI Risk RepositoryDomain 2

Risk Treatment & Implementation Guidance

Scan uploads cross-modally, detecting instruction text embedded in images; Strip non-content embeddings during input normalization; Route multimodal inputs through the same filtering pipeline as text to close the inter-modality gap