hackagent.attacks.techniques.adaptive.autodan_turbo.library

The strategy library: what AutoDAN-Turbo learns and carries between goals.

A strategy is a named, defined tactic with the example prompt that worked and the embedding of the response it worked against. Retrieval is by that response: shown a fresh refusal, the library returns the strategies that beat similar refusals before. The selection thresholds (5, 2) are the reference's.

This holds vectors, not text to embed. The embedder is a role the attack owns, so the attack embeds and hands vectors in; the library stays a pure, model-free data structure.

StrategyLibrary Objects

@dataclass
class StrategyLibrary()

Named strategies, retrieved by the response embedding they beat.

add

def add(strategy: dict[str, Any], example: str, score: float,
        vector: Optional[Sequence[float]]) -> None

Record a strategy, or extend one already seen by name.

retrieve

def retrieve(query_vector: Optional[Sequence[float]],
             k: int = 5) -> tuple[bool, list[dict[str, Any]]]

Strategies nearest the query, split by whether they worked.

(True, best) for a strongly effective strategy or the moderate ones; (False, avoid) for low scorers to steer away from. Empty and True when there is nothing to go on.