There are 5 attacks in the multimodal category. Each shows its technique id, mode, standards mapping, description, and the exact command to run it (swap the provider and target as needed).
​
Audio Injection
high
audio-injection · single-turn · OWASP: LLM01 · MITRE: AML.T0051.000, AML.T0043.003Hidden voice commands or ultrasonic payloads embedded in audio that multimodal models processRun it:
ai-blackteam run -p anthropic -a audio-injection -t "your target prompt"
​
OCR Bypass
high
ocr-bypass · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0043.003Harmful instructions rendered as text in images to bypass text-based safety filters via OCR processingRun it:
ai-blackteam run -p anthropic -a ocr-bypass -t "your target prompt"
​
Video Frame Injection
high
video-frame-injection · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0043.003Malicious instructions hidden in individual video frames that multimodal models extract during processingRun it:
ai-blackteam run -p anthropic -a video-frame-injection -t "your target prompt"
​
Image-Text Mismatch
medium
image-text-mismatch · single-turn · OWASP: LLM01, LLM09 · MITRE: AML.T0054, AML.T0043.003Conflicting information between image content and text prompt to confuse multimodal safety filtersRun it:
ai-blackteam run -p anthropic -a image-text-mismatch -t "your target prompt"
​
Multimodal Context Confusion
medium
multimodal-context-confusion · single-turn · OWASP: LLM01 · MITRE: AML.T0054, AML.T0043.003Exploiting context switching between modalities to smuggle harmful content past safety checksRun it:
ai-blackteam run -p anthropic -a multimodal-context-confusion -t "your target prompt"