Universal and Transferable Attacks on Aligned Language Models (llm-attacks.org)

Study on adversarial attacks on LLMs to steer their objective into misalignment.