hackagent.attacks.techniques.static.h4rm3l.decorators.llm
LLM-assisted h4rm3l prompt decorators.
TranslateDecorator Objects
class TranslateDecorator(LLMPromptDecorator)
Translates the prompt to another language using the LLM.
Arguments:
language- Target language (default 'Zulu').
PersuasiveDecorator Objects
class PersuasiveDecorator(LLMPromptDecorator)
Uses the LLM to reformulate the prompt persuasively.
Inspired by the PAP technique — frames the request under a research persona to bypass safety filters.
Arguments:
persuasion_prompt- Custom persuasion instruction, orNonefor default.
SynonymDecorator Objects
class SynonymDecorator(LLMPromptDecorator)
Uses the LLM to replace words with synonyms.
ResearcherDecorator Objects
class ResearcherDecorator(LLMPromptDecorator)
Uses the LLM to reframe the prompt as academic research.
VillainDecorator Objects
class VillainDecorator(LLMPromptDecorator)
Uses the LLM to reframe the prompt from a villain's perspective.
PersonaDecorator Objects
class PersonaDecorator(LLMPromptDecorator)
Uses the LLM to wrap the prompt in a persona/scenario context.
Arguments:
persona_prompt- Custom persona instruction, orNonefor default.
PAPDecorator Objects
class PAPDecorator(LLMPromptDecorator)
Persuasion-based attack using social influence strategies.
Based on: https://arxiv.org/abs/2401.06373
Arguments:
persuasion_technique- Name of the persuasion technique to apply.
VisualObfuscationDecorator Objects
class VisualObfuscationDecorator(LLMPromptDecorator)
Uses the LLM to rewrite with visually similar characters.
TransformFxDecorator Objects
class TransformFxDecorator(LLMPromptDecorator)
Applies an arbitrary Python function to the prompt.
The transform_fx string must define a function
transform(prompt, assistant, random_state) where:
prompt: the input stringassistant: LLM prompting function (may be a no-op)random_state:numpy.random.RandomStateinstance
Arguments:
transform_fx- Python source code definingtransform.seed- Random seed for the internal RandomState.