What AI says about llm-adaptive-attacks

LLM Adaptive Attacks is a research project that demonstrates how to jailbreak safety-aligned large language models using simple adaptive techniques like random search on adversarial suffixes and prompt prefilling. The project achieves nearly 100% attack success rates on models including GPT-3.5/4, Llama 2, Gemma, and Claude by leveraging logprobs access and model-specific vulnerabilities.