DeepKeep discloses InkJect, an image-hidden prompt attack

Unite.AI reports that DeepKeep uncovered InkJect, an attack technique that hides malicious prompts inside images. The disclosure matters because multimodal AI systems increasingly ingest screenshots, documents, and photos as instructions or context. Attacks that embed instructions in visual inputs expand the prompt-injection problem beyond text fields and chat windows.

Key details: DeepKeep disclosed a technique called InkJect, The attack hides malicious prompts inside images, The target is multimodal AI systems that interpret visual inputs.

Why it matters: As agents consume images and documents, prompt injection becomes a multimodal security problem that ordinary text filters cannot solve alone.