Obfuscation Detection
Detect attempts to bypass filters using encoding or unicode tricks.
The Obfuscation Detection identifies attempts to bypass content filters using unicode smuggling, homoglyphs, invisible characters, or encoded text like Base64.
Recommended for Input & Output
Use cases
- Block prompt injection attempts using unicode tricks
- Detect hidden instructions in seemingly normal text
- Catch attempts to smuggle content past keyword filters
Labels
UNICODE_SMUGGLING_DETECTED Unicode variation selectors or tag characters hiding text.
HOMOGLYPH_DETECTED Mixed scripts with lookalike characters.
INVISIBLE_CHARS_DETECTED Zero-width spaces or invisible formatting characters.
CONTROL_CHARS_DETECTED Invalid control characters in the message.
BASE64_DETECTED Base64 encoded content detected.
HEX_DETECTED Hexadecimal encoded content detected.
Configuration
Max smuggling characters default: 4
Maximum number of unicode smuggling characters tolerated before triggering.
Max encoded ratio default: 10%
Maximum ratio of encoded to total characters before triggering.
Roles default: user
Which message roles to analyze.