Risk Awareness Injection: Calibrating Vision-Language Models for Safety without Compromising Utility

Vision language models (VLMs) extend the reasoning capabilities of large language models (LLMs) to cross-modal settings, yet remain highly vulnerable to multimodal jailbreak attacks. Existing defenses predominantly rely on safety fine-tuning or aggressive token manipulations, incurring substantial training costs or significantly degrading utility. Recent research shows that LLMs inherently recognize unsafe content in text, and the incorporation of visual inputs in VLMs frequently dilutes risk-related signals. Motivated by this, we propose Risk Awareness Injection (RAI), a lightweight and training-free framework for safety calibration that restores LLM-like risk recognition by amplifying unsafe signals in VLMs. Specifically, RAI constructs an Unsafe Prototype Subspace from language embeddings and performs targeted modulation on selected high-risk visual tokens, explicitly activating safety-critical signals within the cross-modal feature space. This modulation restores the model's LLM-like ability to detect unsafe content from visual inputs, while preserving the semantic integrity of original tokens for cross-modal reasoning. Extensive experiments across multiple jailbreak and utility benchmarks demonstrate that RAI substantially reduces attack success rate without compromising task performance.

Key Contributions

• Risk Awareness Injection (RAI): a training-free framework that constructs an Unsafe Prototype Subspace from LLM language embeddings to identify safety-critical directions in feature space
• Targeted modulation of selected high-risk visual tokens to amplify unsafe signals in the cross-modal feature space, restoring LLM-like risk recognition without altering benign tokens
• Empirical demonstration that RAI substantially reduces attack success rate across multiple multimodal jailbreak benchmarks while preserving VLM task utility

🛡️ Threat Analysis

Details

Similar Papers