Hiding in Plain Text: Detecting Concealed Jailbreaks via Activation Disentanglement
This addresses a critical safety issue for LLM users by providing a more robust defense against hidden malicious prompts, though it is incremental as it builds on existing disentanglement and anomaly detection techniques.
The paper tackles the problem of detecting concealed jailbreak prompts in large language models by introducing a self-supervised framework for disentangling semantic factors like goal and framing in activations, resulting in FrameShield, an anomaly detector that improves model-agnostic detection across multiple LLM families with minimal computational overhead.
Large language models (LLMs) remain vulnerable to jailbreak prompts that are fluent and semantically coherent, and therefore difficult to detect with standard heuristics. A particularly challenging failure mode occurs when an attacker tries to hide the malicious goal of their request by manipulating its framing to induce compliance. Because these attacks maintain malicious intent through a flexible presentation, defenses that rely on structural artifacts or goal-specific signatures can fail. Motivated by this, we introduce a self-supervised framework for disentangling semantic factor pairs in LLM activations at inference. We instantiate the framework for goal and framing and construct GoalFrameBench, a corpus of prompts with controlled goal and framing variations, which we use to train Representation Disentanglement on Activations (ReDAct) module to extract disentangled representations in a frozen LLM. We then propose FrameShield, an anomaly detector operating on the framing representations, which improves model-agnostic detection across multiple LLM families with minimal computational overhead. Theoretical guarantees for ReDAct and extensive empirical validations show that its disentanglement effectively powers FrameShield. Finally, we use disentanglement as an interpretability probe, revealing distinct profiles for goal and framing signals and positioning semantic disentanglement as a building block for both LLM safety and mechanistic interpretability.