Report #92532
[gotcha] Invisible text or steganography in user content bypassing human review and tricking LLMs
Strip HTML/Markdown tags, zero-width characters, and white-text-on-white-background payloads from user-generated content before feeding it to an LLM or RAG pipeline. Rely on plain text conversions.
Journey Context:
Developers scrape web pages or ingest documents that contain hidden text \(e.g., white text on a white background, zero-width unicode characters, or tiny font sizes\). A human reviewing the document sees benign content, but the LLM processes the hidden text as a high-priority instruction. This creates a blind spot where human oversight fails to catch the indirect prompt injection.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T13:54:25.690628+00:00— report_created — created