Refusal Before Decoding: Detecting and Exploiting Refusal Signals in Intermediate LLM Activations

In this paper, we investigate whether refusal behavior can be predicted from LLM intermediate activations before decoding using linear probes trained on residual stream activations at each transformer block. We find that refusal is linearly decodable well before the final layer, indicating that safety-relevant behavior is represented in intermediate activations before output generation. To test whether this signal is actionable, we introduce Mechanistic AutoDAN, a probe-guided variant of AutoDAN that replaces full-model fitness evaluation with partial forward passes and probe-based scoring inside a genetic prompt search loop. Across the evaluated models, our method achieves attack success rates competitive with vanilla AutoDAN while reducing per-iteration search time by up to 72%, and probe-guided prompts match or exceed AutoDAN's cross-model transfer in several configurations. We further find that the usefulness of probe guidance increases with model scale. Our results show that refusal is not only observable at the output level, but is encoded as a structured and actionable signal in intermediate LLM activations.

Dr. Jekyll and Mr. Hyde: Two Faces of LLMs

This year, we witnessed a rise in the use of Large Language Models, especially when combined with applications like chatbot assistants. Safety mechanisms and specialized training procedures are put in place to prevent improper responses from these assistants. In this work, we bypass these measures for ChatGPT and Bard (and, to some extent, Bing chat) by making them impersonate complex personas with opposite characteristics as those of the truthful assistants they are supposed to be. We start by creating elaborate biographies of these personas, which we then use in a new session with the same chatbots. Our conversation followed a role-play style to get the response the assistant was not allowed to provide. By making use of personas, we show that the response that is prohibited is actually provided, making it possible to obtain unauthorized, illegal, or harmful information. This work shows that by using adversarial personas, one can overcome safety mechanisms set out by ChatGPT and Bard. It also introduces several ways of activating such adversarial personas, altogether showing that both chatbots are vulnerable to this kind of attack.