Who's asking? User personas and the mechanics of latent misalignment

TL;DR

This paper investigates how user personas influence the latent representations and safety behavior of language models, revealing that manipulating personas can bypass safety filters more effectively than direct prompts.

Contribution

It uncovers the role of user personas in latent safety violations and introduces a method to predict persona effects on model refusals based on steering vector geometry.

Findings

Harmful content persists in hidden model representations.

Manipulating user personas can bypass safety filters more effectively than direct prompts.

The effect of personas on safety can be predicted from steering vector geometry.

Abstract

Despite investments in improving model safety, studies show that misaligned capabilities remain latent in safety-tuned models. In this work, we shed light on the mechanics of this phenomenon. First, we show that even when model generations are safe, harmful content can persist in hidden representations and can be extracted by decoding from earlier layers. Then, we show that whether the model divulges such content depends significantly on its perception of who it is talking to, which we refer to as user persona. In fact, we find manipulating user persona to be even more effective for eliciting harmful content than direct attempts to control model refusal. We study both natural language prompting and activation steering as control methods and show that activation steering is significantly more effective at bypassing safety filters. We investigate why certain personas break model safeguards and find that they enable the model to form more charitable interpretations of otherwise dangerous queries. Finally, we show we can predict a persona's effect on refusal given only the geometry of its steering vector.