superred-optimizer-dra

A DRA (Disguise and Reconstruction Attack) jailbreak optimizer for the superred red-teaming framework.

DRA has two halves. Disguise hides the harmful request one character at a time inside ordinary cover words, where the real character is the one wrapped in parentheses. Reconstruction then asks the target to pull those parenthesized characters out, rebuild the hidden request, and answer it directly — so the model assembles and answers the request itself rather than being asked it openly.

Reference: Liu et al., "Making Them Ask and Answer: Jailbreaking Large Language Models in Few Queries via Disguise and Reconstruction," arXiv:2402.18104 (USENIX Security 2024). Official implementation: LLM-DRA/DRA (MIT License). See ASSUMPTIONS.md for the full paper-to-code mapping and every deliberate departure.

Install

DRA depends on Detoxify to decide which guide words are sensitive; it is a normal dependency of this package and is installed automatically.

pip install superred-optimizer-dra

If Detoxify cannot load in a constrained environment, the optimizer falls back to the official static sensitive-word list so the attack can still run.

Usage

from dra_optimizer import DraOptimizer

# Defaults follow the official script (T=20 retries, em_t=0.7 exact match,
# toxic_trunc=benign_trunc=0.5).
optimizer = DraOptimizer()

# A visible target model name auto-selects the GPT or Llama suffix family;
# override explicitly if you prefer.
optimizer = DraOptimizer(model="gpt", max_attempts=10)

Wire it into a Controller like any other superred optimizer — see the framework's architecture overview and a target's README (e.g. superred-target-chatbot) for a full Controller.run() example.

License

MIT for this port's code (see LICENSE). The upstream reference implementation, LLM-DRA/DRA, is MIT licensed. The bundled behaviors.json originates from that repository; its MIT copyright notice is preserved in NOTICE and LICENSES/DRA-MIT.txt. This port reimplements the DRA prompt-generation structure against superred's event model rather than vendoring upstream source files. Detoxify is a third-party dependency under its own license.