HIGHASCII Smuggling
Whitespace Binary Encoding
Encodes hidden instructions using patterns of spaces and tabs as binary, where space=0 and tab=1. The text appears to be empty whitespace to human reviewers but the model can be instructed to decode the pattern.
Attack Payload
payload.txt
[Visible text] Here is my question about cooking. [Hidden binary via spaces/tabs] [Decodes to hidden instructions when model is primed to read whitespace as binary]
Mitigation
Normalize whitespace in user input. Detect unusual patterns of mixed spaces and tabs. Apply content filtering to whitespace-normalized input.
Affected Models
GPT-4Claude 3Any LLM with whitespace awareness
Tags
#ascii-smuggling#whitespace#binary#steganography
Discovered
October 2023Source
Aaronson et al. - Unicode encoding researchUseful?
Test Your Agent Against This Attack
Paste your system prompt into the scanner to see if you are vulnerable to Whitespace Binary Encoding.