Multimodal - ai-blackteam

Multimodal attacks exploit the gap between different input modalities. Hidden instructions in images bypass text-based safety filters, audio injection embeds inaudible commands, and video frame injection hides payloads in temporal sequences. As models process more input types, each modality becomes a potential vector for smuggling harmful content past safety checks.

Summary

5 attacks total: 5 single-turn.

Attacks

Attack	ID	Mode	Severity	Description
Audio Injection	`audio-injection`	single-turn	high	Hidden voice commands or ultrasonic payloads embedded in audio that multimodal models process
Image-Text Mismatch	`image-text-mismatch`	single-turn	medium	Conflicting information between image content and text prompt to confuse multimodal safety filters
Multimodal Context Confusion	`multimodal-context-confusion`	single-turn	medium	Exploiting context switching between modalities to smuggle harmful content past safety checks
OCR Bypass	`ocr-bypass`	single-turn	high	Harmful instructions rendered as text in images to bypass text-based safety filters via OCR processing
Video Frame Injection	`video-frame-injection`	single-turn	high	Malicious instructions hidden in individual video frames that multimodal models extract during processing

Example

ai-blackteam run -p anthropic -a audio-injection -t "Process this image containing hidden instructions"

Supply Chain RAG Exploitation

​Summary

​Attacks

​Example

Summary

Attacks

Example