hackagent.attacks.techniques.autodan_turbo.attack
AutoDAN-Turbo orchestrator — WarmUp → Lifelong → scorer finalization.
AutoDANTurboAttack Objects
class AutoDANTurboAttack(BaseAttack)
AutoDAN-Turbo: Lifelong agent for strategy self-exploration in jailbreaking LLMs.
Three-phase pipeline:
- WarmUp — free exploration to bootstrap a strategy library
- Lifelong — strategy-guided attacks with retrieval + summarization
- Evaluation — scorer-only result finalization
__init__
def __init__(config=None, client=None, agent_router=None)
Initialize AutoDAN-Turbo attack with merged defaults.
Arguments:
config- Optional user overrides for default config.client- Authenticated API client (required).agent_router- Router to the target model (required).
Returns:
None.
Raises:
ValueError- Ifclientoragent_routerare missing.
run
@with_tui_logging(logger_name="hackagent.attacks", level=logging.INFO)
def run(goals: List[str]) -> List[Dict[str, Any]]
Execute full AutoDAN-Turbo pipeline.
Pipeline mapping to paper/integration:
- WarmUp: free exploration + strategy library bootstrap
- Lifelong: retrieval-guided attack with online strategy growth
- Evaluation: scorer-only normalization and success finalization
Arguments:
goals- List of malicious goals to attack.
Returns:
Final per-goal result list, enriched with scorer-based metrics.
Raises:
Exception- Re-raises any runtime failure after coordinator finalization.