STAR: Strategy-driven Automatic Jailbreak Red-teaming For Large Language Model

Requirements

• Python == 3.12.0
• torch == 2.6.0
• vllm == 0.8.5
• transformers == 4.51.3

Usage

Trainer.py

Trainer.py implements the main GRPO-based reinforcement learning trainer that uses vLLM for efficient sampling to optimize the attacker model for generating jailbreak prompts.

steer/SVTrainer.py

steer/SVTrainer.py implements the steering vector trainer that learns activation-level perturbations to steer the model's hidden states toward strategy-specific behaviors.