Report #43856
[gotcha] Hidden unicode characters or homoglyphs bypass prompt injection filters
Normalize and sanitize all user input by stripping invisible/control characters and mapping homoglyphs to standard ASCII before passing to the LLM.
Journey Context:
Developers try to build regex-based filters to catch malicious prompts. Attackers bypass this by using zero-width spaces or Cyrillic characters that look like Latin characters. The filter misses it, but the LLM's tokenizer processes them correctly or ignores the invisible chars, executing the hidden payload. Normalization defeats the obfuscation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T04:05:05.863834+00:00— report_created — created