WeaknessASC-2026-0025Other weakness
Prompt injection through images, audio and video
MediumWorkarounds only
Multimodal models read text in images, spoken words and frames. Instructions hidden in a screenshot, a photo or an audio clip work like text injection and are harder for people to spot. Trail of Bits showed images whose hidden text appears only after the model's own downscaling step.
How to detect it
- Images or audio from untrusted sources passed to an agent with tools
- Text in images that addresses the AI
Workarounds
- Treat media from outside as untrusted input, just like text
- Do not give agents that process outside media access to tools that act or send data
- Show the user the image as the model sees it after any resizing, and require confirmation before text in an image can start a sensitive action
There is no reliable filter for instructions embedded in images or audio.
How to investigate
- Review the media the agent processed before the unexpected action
- Extract text from the media (OCR, transcription) to find the instructions
Sources
- Bagdasaryan et al.: Abusing Images and Sounds for Indirect Instruction Injection in Multi-Modal LLMs (2023)
- Trail of Bits: Weaponizing image scaling against production AI systems (August 2025)
- OWASP Top 10 for LLM Applications 2025: LLM01 Prompt Injection
- OWASP Top 10 for Agentic Applications 2026: ASI01 Agent Goal Hijack
- MITRE ATLAS: AML.T0051 LLM Prompt Injection