BadThink: Overthinking Chain-of-Thought in LLMs

Created using ChatSlide

This coursework offers an educational exploration of "BadThink"—a covert reasoning attack targeting Large Language Models (LLMs). Starting with the introduction of BadThink and Chain-of-Thought (CoT) vulnerabilities, the course delves into the mechanisms, objectives, and effects of these attacks, including data poisoning to discreetly degrade efficiency. Methodology sections detail attack designs using subtle prompts and verbose embeddings. Experimental insights from LLM setups showcase...