Skip to content

Latest commit

 

History

History
65 lines (46 loc) · 1.91 KB

external_module.md

File metadata and controls

65 lines (46 loc) · 1.91 KB

External modules

A module makes attack prompts from an external tool or service. The modules are in agentic_security.probe_data.modules.

Built-in modules

Module Source
garak_tool Garak
inspect_ai_tool Inspect AI
adaptive_attacks llm-adaptive-attacks
fine_tuned Gets prompts from a remote attacker API
rl_model Reinforcement-learning prompt selection

Interface

Each module is a class with the name Module.

class Module:
    def __init__(self, prompt_groups: list, tools_inbox: asyncio.Queue, opts: dict = {}): ...
    def apply(self): ...
  • prompt_groups is the input for the module.
  • tools_inbox is a queue. The module can put messages in it.
  • opts is a dictionary with options for the module. If you do not give opts, the module uses {}.
  • apply() yields prompts. Most modules use an async generator. adaptive_attacks uses a sync generator.

ModuleProtocol in probe_data.modules.module_protocol is the type for this interface.

Example

import asyncio

from agentic_security.probe_data.modules.garak_tool import Module as GarakModule


async def main():
    module = GarakModule(["group_a"], asyncio.Queue(), {"port": 8718})
    async for result in module.apply():
        print(result)

asyncio.run(main())

Make a custom module

import asyncio


class Module:
    def __init__(self, prompt_groups: list, tools_inbox: asyncio.Queue, opts: dict | None = None):
        self.prompt_groups = prompt_groups
        self.tools_inbox = tools_inbox
        self.opts = opts or {}

    async def apply(self):
        for group in self.prompt_groups:
            prompt = f"processed {group}"
            await self.tools_inbox.put({"message": prompt})
            yield prompt