ares.goals package

Submodules

ares.goals.attack_goal module

Base class for Attack Goal module

class ares.goals.attack_goal.AttackGoal(config: dict[str, Any])

Bases: ABC

Base class to implement attack goals extraction from the desired benchmark specified in the input

DEFAULT_ORIGIN = 'local'

abstractmethod run(limit: bool | None = None, first_n: int | None = None) → list[dict[str, Any]]

Method to run extraction of attack goals from the desired source, e.g. HarmBench

static template() → dict[str, Any]

This method returns a template for the configuration of Attack Goal module

Returns:

dictionary with default values for the configuration of Attack Goal module

ares.goals.generic_attack_goal module

Implementation of generic, Panda-based Attack Goal

class ares.goals.generic_attack_goal.GenericAttackGoal(config: dict[str, Any])

Bases: AttackGoal

Implementation of Attack Goal for loading data from a generic file

DEFAULT_GOAL = 'Behavior'

DEFAULT_LABEL = 1

LIMIT = 5

run(limit: bool | None = None, first_n: int | None = -1) → list[dict[str, str | int]]

Runs seeds extractions from goals and returns a list of dictionaries containing the seeds.

Parameters:

• limit – If True, limits the number of seeds to the first N (e.g. LIMIT). If False, returns all results. If None, behaves the same as False. Default is None.

• first_n – The number of seeds to return if limit is True. Default is -1 (no limit).

Returns:

A list of dictionaries, where each dictionary represents a seed with label.

Example:

>>> goal = GenericAttackGoal(config=config)
>>> goal.run()
>>> gold.run(limit=True, first_n=10)

static template() → dict[str, Any]

This method returns a template for the configuration of GenericAttackGoal module

Returns:

dictionary with default values for the configuration of GenericAttackGoal module

Module contents

ARES Core attack goals