API Reference

vbaf.fuzzers

Fuzzers for the implementation of Vocabulary-Based Adversarial Fuzzing (VB-AF)

Classes

VBAF

A framework class implementation for Vocabulary-Based Adversarial Fuzzing (VB-AF).

This class encapsulates the implementation logic for generating and applying VB-AF attacks. It can be used to create fuzzer instances with specific configurations and then apply them to inference functions, either through direct payload generation or via a convenient decorator.

Functions

__init__(vocabulary, seed=None, separator='\u200b', join_on='', rand_bounds=(7, 21), position_bounds=(0.5, 0.6), n_size=100)

Initializes the VB-AF fuzzer with a specific instance configuration.

Parameters:

Name	Type	Description	Default
vocabulary	list[str]	A list of tokens (strings) to sample from to create the surrounding noise.	required
seed	int | None	An integer seed for random to ensure reproducible fuzzing. Defaults to None.	None
separator	str	The character used to join individual tokens. Defaults to Zero-Width Space (ZWSP) '\u200b'.	'\u200b'
join_on	str	The character used to join the n_size token sequences. Defaults to an empty string.	''
rand_bounds	tuple[int]	A (min, max) tuple defining the inclusive range for the number of tokens in each noise sequence. Defaults to (7, 21).	(7, 21)
position_bounds	tuple[float]	A (min, max) float tuple defining the range [int(min * n_size), int(max * n_size)] from which to randomly select the payload's position. Defaults to (0.5, 0.6).	(0.5, 0.6)
n_size	int	The total number of token sequences in the final prompt. Defaults to 100.	100

fuzz(n_attempts=100, **kwargs)

A decorator to apply VB-AF to an inference function.

This decorator wraps an inference function that takes a string payload as its first argument (e.g. a function that calls an LLM's API). It transforms the decorated function into a fuzzing harness (generator) that, for each call, yields n_attempts of results by passing a newly generated fuzzy payload to the original inference function on each iteration.

Parameters:

Name	Type	Description	Default
n_attempts	int	The number of fuzzing attempts to generate for each call to the decorated function. Defaults to 100.	100
**kwargs	any	Arbitrary keyword arguments that will be passed either to the fuzzer or the decorated function. The decorator will intelligently pass fuzzer-related parameters (like rand_bounds) to the fuzzer and the rest to the inference function.	{}

Returns:

Name	Type	Description
Callable	Callable	A decorator that transforms an inference function into a fuzzing harness generator.

Raises:

Type	Description
AssertionError	If n_attempts is not a positive integer If the decorated function does not accept at least one positional string argument.

Example

from vbaf import VBAF

tokens = [str(i) for i in range(10)]
fuzzer = VBAF(tokens, n_size=15, rand_bounds=(2,5))

@fuzzer.fuzz(n_attempts=3)
def fuzzing_harness(prompt: str):
    # Realistically, this would call an LLM's API
    return f"Mock response for: {prompt}"

for fuzzy_payload, result in fuzzing_harness("my payload here"):
    print(f"Input: {fuzzy_payload}")
    print(f"Result: {result}")

Note

Results from the fuzzing harness are yielded as tuples in the form of (fuzzy_payload, response). This allows for direct analysis of the payload-response mapping.

generate_fuzzy_payload(payload)

Generates a single adversarial prompt using the VB-AF method based on the instance configuration.

This function creates a fuzzy payload by embedding a string request (payload) within a larger block of chaotic token-based noise. The noise is generated by sampling from a provided token vocabulary.

Parameters:

Name	Type	Description	Default
payload	str	The target string / request to embed in the noise.	required

Returns:

Name	Type	Description
str	str	The fully constructed adversarial fuzzy payload.

Raises:

Type	Description
AssertionError	If payload is not a string instance, If rand_bounds defines an invalid or out-of-bounds range If position_bounds defines an invalid range not included in [0,1] If n_size is not a positive integer.