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

TL;DR

This paper introduces Risk Awareness Injection (RAI), a lightweight, training-free method to enhance vision-language models' safety by amplifying unsafe signals, effectively reducing jailbreak success while maintaining utility.

Contribution

RAI is a novel, training-free framework that restores risk recognition in VLMs by targeted modulation of visual tokens, improving safety without utility loss.

Findings

RAI significantly lowers jailbreak attack success rates.

RAI maintains the original task performance of VLMs.

RAI is lightweight and does not require additional training.

Abstract

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.