Joshua Judges Ruth
A practitioner testing an LLM-powered financial-document reader describes hiding misleading text and instructions in PDFs, then building separate checks for suspicious differences between the document’s text, accessibility metadata and rendered pages.
The author describes limitations in each check, including accessibility-text substitutions that can evade two of them. This is work in progress, not proof that prompt injection is solved; the author has not encountered an attack in the wild. The Lobsters discussion had no comments when checked.