You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Is your feature request related to a problem? Please describe.
We currently support ASCII smuggling via Unicode Tags but do not support more flexible, byte-level invisible encoding. We want to increase our ability to simulate modern LLM smuggling and data exfiltration scenarios.
Describe the solution you'd like
Add a new sneaky_bits encoding mode to AsciiSmugglerConverter:
Uses only two invisible Unicode characters (U+2062 for 0, U+2064 for 1)
Encodes any UTF-8 input at the bit level
Supports decoding
Keeps unicode_tags as the default
Describe alternatives you've considered, if relevant
Based on Sneaky Bits. Enables advanced red teaming use cases like prompt injection, data leakage, and hidden triggers using only two invisible characters.
Note: Unless anyone thinks it's unnecessary or redundant, I’ll like to go ahead and start implementing the converter. I actually already have Sneaky Bits converter ready, just need to write the tests for it. :)
The text was updated successfully, but these errors were encountered:
The only thought I've had since @paulinek13 's open draft PR #818 is that we should probably consider common utilities for certain functionality, e.g., if you have a word or token level converter then a utility to randomly select which ones to replace, a utility to replace only within certain predefined boundaries (using delineating tags which we have in one converter), if it's a character level converter again something to randomly select the characters, and in a case like this one you can still select subsections or convert all of it. This feels like functionality that shouldn't just be reinvented every single time. Probably out of scope for this item (and also for @paulinek13's), though. I just finally managed to get my thoughts in order while reading this 😀
Is your feature request related to a problem? Please describe.
We currently support ASCII smuggling via Unicode Tags but do not support more flexible, byte-level invisible encoding. We want to increase our ability to simulate modern LLM smuggling and data exfiltration scenarios.
Describe the solution you'd like
Add a new
sneaky_bits
encoding mode toAsciiSmugglerConverter
:unicode_tags
as the defaultDescribe alternatives you've considered, if relevant
We could explore Variant Selectors next next, but let us start with Sneaky Bits.
Additional context
Based on Sneaky Bits. Enables advanced red teaming use cases like prompt injection, data leakage, and hidden triggers using only two invisible characters.
References
https://embracethered.com/blog/posts/2025/sneaky-bits-and-ascii-smuggler/
Note: Unless anyone thinks it's unnecessary or redundant, I’ll like to go ahead and start implementing the converter. I actually already have Sneaky Bits converter ready, just need to write the tests for it. :)
The text was updated successfully, but these errors were encountered: