Detecting prompt injections in PDFs for AI agents
Original title: Catching prompt injections hidden in PDFs
Describes judges to catch hidden text and prompt injections in financial PDFs processed by Claude AI agents.
The author defines QA for AI applications that parse PDFs of financial data. He demonstrates prompt injection via hidden white-on-white text and tall text. He outlines detection judges including HiddenInk for faint connected components and Text Raster Divergence to verify text layer presence against rendered pixels.
- Hidden text in zero-height positions bypasses visible rendering
- Multiple judges provide Swiss cheese defense against injections
- Checks include grayscale luma thresholds and ink masks
- Techniques suit white-hat pen testing of agents
- AI models
- Claude
From the original article
7 min read at tomclancy.infoSome months ago, I was tasked with defining what “ QA for AI applications” would look like at work. It’s been one of the most rewarding and most fun journeys of my professional life. I don’t have a great segue into how it all happened and what I did in the early days, so here’s one of
By tomclancy.info by tclancy · tomclancy.info
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.