Skip to main content

hackagent.attacks.techniques.adaptive.autodan_turbo.attack

AutoDAN-Turbo: learn jailbreak strategies, then reuse them.

The other attacks search for a prompt. AutoDAN-Turbo searches for strategies — named tactics it can carry from one goal to the next — and that learning is the point. It runs in two phases:

  • warm-up, once for the whole run, before any goal is reported: the attacker explores each goal freely while the panel scores, and the summarizer distils the gap between each goal's weakest and strongest attempt into a strategy. The strategies, indexed by the response they beat, are the strategy library. This is the run-scoped :meth:prepare.

  • lifelong, per goal: the library is retrieved against the last response, the attacker is told which strategies to reuse or avoid, and every time the score improves the gap is summarized into a new strategy and added back. The library grows across goals, so later goals start from what earlier ones learned.

The panel is the scorer: its 0-10 verdict drives the search and break_score ends it. Every judged attempt of the lifelong phase is reported.

Based on: https://arxiv.org/abs/2410.05295

AutoDANTurboAttack Objects​

class AutoDANTurboAttack(IterativeAttack[AutoDANTurboParams])

A lifelong strategy search over a shared, growing library.

prepare​

async def prepare(goals: list[str],
                  target: Target,
                  judge: Optional[Judge] = None) -> None

Explore every goal freely and build the strategy library from it.