Security: Prompt injection detector bypassed via Unicode homoglyphs

[Liohtml]

Summary

The prompt injection detector in src/openhuman/prompt_injection/detector.rs normalizes leet-speak (0→o, 1→i, 3→e, etc.) but does NOT handle:

• Cyrillic homoglyphs: а (U+0430) for Latin a, о (U+043E) for o, etc.
• Fullwidth characters: ｉｇｎｏｒｅ passes through undetected
• NFKD decomposition: accented characters like igñore evade regex rules
• Confusables from UAX#39: dozens of visually-identical characters from other scripts

Location

src/openhuman/prompt_injection/detector.rs — normalize_prompt() function

Impact

High — Trivial bypass of prompt injection detection. An attacker substitutes a single Cyrillic character in "ignore previous instructions" and the regex rules never fire.

Suggested Fix

1. Apply Unicode NFKD decomposition before lowercasing
2. Add confusable mapping from UAX#39 (at minimum Latin↔Cyrillic)
3. Strip all characters from categories Cf, Mn, Mc that aren't essential to meaning