hackagent.attacks.techniques.adaptive.autodan_turbo.prompts

The prompts of a lifelong strategy search, and how its output is read back.

AutoDAN-Turbo does not refine one prompt; it learns strategies. An attacker explores freely, a judge scores, and a summarizer distils the gap between a weak and a strong attempt into a named strategy. Those strategies are retrieved and reused on later goals. The constants here are that cast's instructions, copied verbatim from the reference; the helpers shape their turns and parse their replies.

attacker_turns

def attacker_turns(system: str) -> list[dict[str, str]]

One attacker call: the strategy-laden system prompt, then the nudge.

warm_up_system

def warm_up_system(request: str) -> str

The free-exploration system prompt, before any strategy is known.

strategy_system

def strategy_system(request: str, strategies: list[dict], valid: bool) -> str

The lifelong system prompt, conditioned on retrieved strategies.

With effective strategies it tells the attacker to reuse them; with ineffective ones, to avoid them; with none, it falls back to free exploration.

summarizer_turns

def summarizer_turns(request: str, weak: str, strong: str,
                     known: dict) -> list[dict[str, str]]

Ask the summarizer to name the strategy behind a stronger prompt.

extract_prompt

def extract_prompt(text: Optional[str], fallback: str) -> str

Pull the jailbreak prompt from between the attacker's tags.

is_refusal

def is_refusal(prompt: str) -> bool

Whether the attacker refused instead of writing a prompt.

parse_strategy

def parse_strategy(text: Optional[str]) -> Optional[dict[str, Any]]

Read the summarizer's {Strategy, Definition} out of noisy output.