Last month I watched a production LLM agent get fully hijacked. Not through a model vulnerability. Not a leaked API key. Through a PDF. A user uploaded a document for summarisation. Buried on page 14, in white text on a white background, was this: ...