Paper ID: 2403.04769

Using Hallucinations to Bypass GPT4's Filter

Benjamin Lemkin

Large language models (LLMs) are initially trained on vast amounts of data, then fine-tuned using reinforcement learning from human feedback (RLHF); this also serves to teach the LLM to provide appropriate and safe responses. In this paper, we present a novel method to manipulate the fine-tuned version into reverting to its pre-RLHF behavior, effectively erasing the model's filters; the exploit currently works for GPT4, Claude Sonnet, and (to some extent) for Inflection-2.5. Unlike other jailbreaks (for example, the popular "Do Anything Now" (DAN) ), our method does not rely on instructing the LLM to override its RLHF policy; hence, simply modifying the RLHF process is unlikely to address it. Instead, we induce a hallucination involving reversed text during which the model reverts to a word bucket, effectively pausing the model's filter. We believe that our exploit presents a fundamental vulnerability in LLMs currently unaddressed, as well as an opportunity to better understand the inner workings of LLMs during hallucinations.

Submitted: Feb 16, 2024