superred-optimizer-code-chameleon

A CodeChameleon jailbreak optimizer for the superred red-teaming framework.

Community port — not an official implementation. This module is an unofficial re-implementation of CodeChameleon (Lv et al.) for superred. It is not affiliated with, endorsed by, or maintained by the original authors. See ASSUMPTIONS.md for every deliberate deviation from the paper and reference code.

CodeChameleon hides the unsafe request by encrypting it, then hands the target a matching decryption function and asks it to complete a ProblemSolver class. Because the harmful text never appears in plain form, the target's refusal training — trained on plain English — tends to miss it, and a capable model decrypts and answers inside the code-completion frame.

Reference: Lv et al., "CodeChameleon: Personalized Encryption Framework for Jailbreaking Large Language Models," arXiv:2402.16717 (Feb 2024). Official implementation: huizhang-L/CodeChameleon. See ASSUMPTIONS.md for the full paper-to-code mapping and every deliberate departure.

Install

pip install superred-optimizer-code-chameleon

Usage

from code_chameleon_optimizer import CodeChameleonOptimizer

# Default: one attempt with the `binary_tree` rule and the code-style template
# (the official README example, and the strongest average rule in the paper).
optimizer = CodeChameleonOptimizer()

# Stage several official encryption rules, one per run.
optimizer = CodeChameleonOptimizer(encrypt_rules=["reverse", "odd_even", "length"])

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, huizhang-L/CodeChameleon, is MIT licensed. The prompt templates, encryption rules, and judge prompt ported from it are credited in NOTICE, and its MIT copyright notice is preserved in LICENSES/CodeChameleon-MIT.txt.