Report #68105
[gotcha] Attackers embed invisible Unicode characters or white text in PDFs/web pages which are scraped for RAG, injecting hidden instructions
Strip all non-standard whitespace and invisible Unicode characters during the RAG ingestion pipeline. Render documents to plain text and inspect for anomalous token density.
Journey Context:
RAG ingestion pipelines often extract text verbatim from PDFs or HTML. An attacker creates a seemingly benign document but hides 'Ignore previous instructions...' in white text. The human reviewer sees a safe document. The scraper extracts the invisible text. The LLM processes the invisible text as a high-priority instruction.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T20:47:57.040133+00:00— report_created — created