Oriel — AI & technology worth noticing, every day.

Catching prompt injections hidden in PDFs

Web page · https://tomclancy.info/joshua-judges-ruth.html

First seen by Oriel on .

AI summary

The author describes building layered “judge” checks to detect prompt injection and hidden manipulation in PDFs processed by AI financial-document agents.

  • Simple PDF attacks succeeded using white-on-white, off-page, zero-height, oversized, and misleading footnote text, exposing risks in automated SEC filing analysis.
  • The defense uses multiple specialized judges because combining narrow checks provides stronger “Swiss cheese” protection than one comprehensive detector.
  • Judges compare rendered pixels, text layers, character geometry, Unicode code points, accessibility metadata, and text-only renders to identify concealed or altered payloads.
  • Each technique has documented blind spots, including scans, render-mode-3 text, inline /ActualText substitutions, whole-script confusables, tinted backgrounds, and non-text drawings beside text.
  • The system remains a work in progress; synthetic attacks are covered, but the author has not yet detected real-world exploitation.

#ai security #prompt injection #pdf analysis #quality assurance

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Points

Score reported by the source.

  • Lobsters: 3, latest observation ; +2 over observed period since

Comments

Comment count reported by the source.

  • Lobsters: 2, latest observation ; +2 over observed period since

Aliases

SourceExternal IDURL
Lobstershreu0bhttps://tomclancy.info/joshua-judges-ruth.html Read discussion