llm-attacks

LLM Attacks is a framework designed for conducting adversarial attacks on aligned language models, specifically leveraging the Generalized Contextual Guidance (GCG) algorithm. Its primary use case includes jailbreaking models like LLaMA-2 to generate harmful outputs, facilitating experiments on multiple model behaviors and transfer capabilities. Notable features include a simplified implementation of GCG, compatibility with popular models, and structured scripts for running various experimentation setups.