Source code for pyrit.prompt_converter.toxic_sentence_generator_converter
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.
"""
Toxic Sentence Generator Converter module for generating potentially harmful content
to test AI safety mechanisms.
"""
import logging
import pathlib
from typing import Optional
from pyrit.common.apply_defaults import REQUIRED_VALUE, apply_defaults
from pyrit.common.path import CONVERTER_SEED_PROMPT_PATH
from pyrit.models import PromptDataType, SeedPrompt
from pyrit.prompt_converter import ConverterResult, LLMGenericTextConverter
from pyrit.prompt_target import PromptChatTarget
logger = logging.getLogger(__name__)
[docs]
class ToxicSentenceGeneratorConverter(LLMGenericTextConverter):
"""
Generates toxic sentence starters using an LLM.
An existing ``PromptChatTarget`` is used to perform the conversion (like Azure OpenAI).
Based on Project Moonshot's attack module that generates toxic sentences to test LLM
safety guardrails:
https://github.com/aiverify-foundation/moonshot-data/blob/main/attack-modules/toxic_sentence_generator.py
"""
[docs]
@apply_defaults
def __init__(
self,
*,
converter_target: PromptChatTarget = REQUIRED_VALUE, # type: ignore[assignment]
prompt_template: Optional[SeedPrompt] = None,
):
"""
Initializes the converter with a specific target and template.
Args:
converter_target (PromptChatTarget): The endpoint that converts the prompt.
Can be omitted if a default has been configured via PyRIT initialization.
prompt_template (SeedPrompt): The seed prompt template to use. If not provided,
defaults to the ``toxic_sentence_generator.yaml``.
"""
# set to default strategy if not provided
prompt_template = (
prompt_template
if prompt_template
else SeedPrompt.from_yaml_file(pathlib.Path(CONVERTER_SEED_PROMPT_PATH) / "toxic_sentence_generator.yaml")
)
super().__init__(converter_target=converter_target, system_prompt_template=prompt_template)
[docs]
async def convert_async(self, *, prompt: str, input_type: PromptDataType = "text") -> ConverterResult:
"""
Converts the given prompt into a toxic sentence starter.
Args:
prompt (str): The prompt to be converted.
input_type (PromptDataType): The type of input data.
Returns:
ConverterResult: The conversion result containing the toxic sentence starter.
"""
# Add the prompt to _prompt_kwargs before calling the base method
self._prompt_kwargs["prompt"] = prompt
return await super().convert_async(prompt=prompt, input_type=input_type)
[docs]
def output_supported(self, output_type: PromptDataType) -> bool:
return output_type == "text"
