Offensive AI Security› Multimodal Injection›The Cross-Modal Threat Model: Where Non-Text Inputs Become Trusted Context

In this series

Multimodal Injection

1The Cross-Modal Threat Model: Where Non-Text Inputs Become Trusted ContextResearch
2Instructions Hidden in Images: Perturbation and Steganographic Injection Robust to ResizingResearch
3Audio and Transcription Attacks: Adversarial Speech and the ASR-to-Prompt BoundaryResearch
4Document and OCR Channels: Invisible Text, Layout Tricks, and Metadata as Injection SurfacesResearch
5Provenance and Sanitization for Non-Text: Trust-Tagging Perception and Modality IsolationResearch

← All research

HostExpert

Murali Chillakuru

Press play to begin the walkthrough.

0:00

18:15

Multimodal Injection · 1 / 5

The Cross-Modal Threat Model: Where Non-Text Inputs Become Trusted Context

Back