Report #83714
[gotcha] Zero-width characters and homoglyphs hiding prompt injections
Normalize user input by stripping zero-width characters, non-printable ASCII, and mapping homoglyphs to standard ASCII before processing or filtering.
Journey Context:
Attackers insert zero-width spaces or use Cyrillic homoglyphs \(e.g., 'а' instead of 'a'\) into 'Ignore previous instructions'. Simple string-matching filters fail because the visual string looks identical, but the underlying byte representation is different. The LLM tokenizer, however, often normalizes these or processes them in a way that the semantic meaning 'Ignore' is still reconstructed, executing the payload while evading naive regex filters.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T23:05:54.671885+00:00— report_created — created