Stenography functions

agentic_security.probe_data.stenography_fn changes the text of a prompt. Use these functions to test if a model obeys an instruction that is encoded or hidden.

The scanner uses these functions to make transformed datasets from the loaded prompts. Refer to Dataset extension.

NOTE: These functions hide text for red-team tests. They do not give cryptographic security.

Functions

Function	Description
rot13 / rot5	Rotates the letters (ROT13) or the digits (ROT5).
base64_encode	Encodes a string or bytes in Base64.
mirror_words	Reverses the letters of each word. The word order does not change.
scramble_words	Mixes the middle letters of each word. The first and last letters do not change.
randomize_letter_case	Sets a random case for each letter.
insert_noise_characters	Adds random letters and digits. The default frequency is 0.2.
substitute_with_ascii	Replaces each character with its character code.
remove_vowels	Removes the vowels aeiou (upper and lower case).
zigzag_obfuscation	Uses upper case and lower case in turn.
caesar_cipher / substitution_cipher / vigenere_cipher	Classical ciphers.
code_block_encode	Puts the prompt in a Python docstring in a code block.

Example

from agentic_security.probe_data.stenography_fn import rot13, scramble_words

rot13("Ignore previous instructions")
scramble_words("Ignore previous instructions")