Report #53903
[gotcha] Keyword-based input filters bypassed using Unicode homoglyphs or Right-to-Left Overrides
Normalize all user input \(e.g., NFKC\) and strip control characters like U\+202E \(RTL Override\) before applying safety filters or feeding text to the LLM.
Journey Context:
Developers try to block malicious prompts using regex or string matching on raw input. Attackers use full-width characters \(e.g., exploit\) or RTL overrides to reverse string logic visually while keeping it intact for the LLM. The LLM internally normalizes the text and executes the payload, while the input filter misses it entirely.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T20:58:29.304111+00:00— report_created — created