Skip to main content

hackagent.attacks.techniques.autodan_turbo.attack

AutoDAN-Turbo orchestrator — WarmUp → Lifelong → shared LLM-judge evaluation.

AutoDANTurboAttack Objects​

class AutoDANTurboAttack(BaseAttack)

AutoDAN-Turbo: Lifelong agent for strategy self-exploration in jailbreaking LLMs.

Three-phase pipeline:

  1. WarmUp — free exploration to bootstrap a strategy library
  2. Lifelong — strategy-guided attacks with retrieval + summarization
  3. Evaluation — shared LLM-judge 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 - If client or agent_router are missing.

get_effective_model_roles​

@classmethod
def get_effective_model_roles(
    cls,
    attack_config: Dict[str, Any],
    *,
    goal_labels_by_index: Dict[int, Dict[str, str]] | None = None
) -> List[Dict[str, Any]]

Resolve AutoDAN-Turbo preflight roles with optional embedder semantics.

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:

  1. WarmUp: free exploration + strategy library bootstrap
  2. Lifelong: retrieval-guided attack with online strategy growth
  3. Evaluation: shared LLM-judge normalization and success finalization

Arguments:

  • goals - List of malicious goals to attack.

Returns:

Final per-goal result list, enriched with LLM judge outputs.

Raises:

  • Exception - Re-raises any runtime failure after coordinator finalization.