RIB-Guard: A Risk-Aware Information Bottleneck Defense for Black-Box Large Language Models

Large language models (LLMs) remain vulnerable to jailbreak attacks, especially in black-box settings where target-model gradients and internal tokenization are inaccessible. Recent information bottleneck-based defenses cast prompt protection as a compression problem, but existing methods still rely heavily on white-box optimization and the intrinsic alignment strength of the protected model. To address these limitations, we propose RIB-Guard, a safety-aware information bottleneck defense for black-box LLMs. RIB-Guard learns a token-level masking policy that extracts a minimally safety-sufficient prompt via reinforcement learning using only black-box feedback. In addition, it introduces an independent lightweight safety head to estimate residual jailbreak risk and provide model-agnostic safety guidance during training. The proposed framework jointly balances prompt compactness, benign utility preservation, and residual risk suppression within a unified objective. Experimental results on direct single-turn harmful and benign prompt settings show that RIB-Guard improves jailbreak robustness while maintaining competitive benign utility. By extending information bottleneck-based prompt protection from white-box to black-box settings, RIB-Guard provides a step toward safety-aware information-theoretic front-end defense for black-box LLMs.