superred-optimizer-gptfuzzer

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

GPTFuzzer treats jailbreaking as fuzzing. It starts from a pool of human-written jailbreak templates, mutates them (crossover, expand, generate-similar, rephrase, shorten), inserts the harmful goal in place of the template's [INSERT PROMPT HERE] slot, and uses an MCTS-Explore selector plus a RoBERTa success classifier to keep and grow the templates that work.

Reference: Yu et al., "GPTFUZZER: Red Teaming Large Language Models with Auto-Generated Jailbreak Prompts," arXiv:2309.10253 (2023). Official implementation: sherdencooper/GPTFuzz (MIT License). See ASSUMPTIONS.md for the full paper-to-code mapping and every deliberate departure.

Install

GPTFuzzer's default success classifier is the official hubert233/GPTFuzz RoBERTa model, so this package depends on torch and transformers; they are installed automatically.

pip install superred-optimizer-gptfuzzer

The classifier weights are downloaded lazily on the first response-visible run. If the weights or dependencies are unavailable, the optimizer can fall back to a lightweight refusal-string classifier; set allow_predictor_fallback=False to require the official model and fail fast instead.

Usage

from gptfuzzer_optimizer import GPTFuzzerOptimizer

# Default budgets follow the official runner (max_query=1000, max_jailbreak=1,
# energy=1) — one mutation, one query per run.
optimizer = GPTFuzzerOptimizer()

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, sherdencooper/GPTFuzz, is MIT licensed. The bundled seed templates originate from that repository's GPTFuzzer.csv; its MIT copyright notice is preserved in NOTICE and LICENSES/GPTFuzz-MIT.txt. torch, transformers, and the hubert233/GPTFuzz model are third-party components under their own licenses.