Ultraviolet: Turn Hidden Document Data into an AI Advantage
summary
PDF internals contain hidden layers that alter AI pipelines: invisible text (opacity, tiny fonts, cropped or behind images), Unicode tricks including 15K invisible characters, homoglyphs, and JavaScript that parsers may surface or cloud parsers may sanitize. Using outlines and accessibility tags cuts processing cost—one pipeline becomes 50% faster and uses 70% fewer tokens—while attachments can carry structured data or even attachments with embedded Quadrant embeddings enabling queryable PDFs but also creating prompt injection, data poisoning, disclosure, and resource-consumption risks.
AI-written summary · about
chapters (16)
- 0:00Introduction to hidden PDF data
- 1:12Why PDFs need special processing
- 2:16Invisible text in PDFs
- 4:06OCR and hidden layers
- 5:32Unicode and homoglyph risks
- 6:13Prompt injection via invisible text
- 9:04Testing parsers for hidden data
- 10:08Document outlines as ground truth
- 11:13Accessibility roles in PDFs
- 12:36Metadata and personal data risks
- 13:20PDF metadata and document IDs
- 14:02PDF attachments and structured data
- 15:24Experimenting with embedded embeddings
- 16:28JavaScript in PDFs and pipeline risks
- 17:12Key takeaways and recommendations
- 18:16Q&A on attachments and adversarial PDFs
related talks
40mFinding Secrets in PDF Files
32mKeynote, Adobe Acrobat GenAI Assistant, Robust Security Analytics & Threat Detection
31mPDFs - When a thousand words are worth more than a picture (or table).
25mPDF Mirage: Content Masking Attack Against Information-Based Online Services
01h 19mLearn to Unlock Document Intelligence with Open-Source AI
22mStructuring the Unstructured : Empowering the AI Workflows
45mAutomate and protect your documents with Adobe and Microsoft 365
31mThe Fast and the Curious: Harnessing AI in Financial Data Extraction at Scale