AI Sec Watch: A Security Intelligence Platform for AI Systems

Luu, T.J.

SafeSteer: Adaptive Subspace Steering for Efficient Jailbreak Defense in Vision Language Models

inforesearchPeer-Reviewed

safetyresearch

Source: IEEE Xplore (Security & AI Journals)July 17, 2026

Summary

Vision Language Models (VLMs, which are AIs that process both images and text) are vulnerable to jailbreak attacks (techniques that trick an AI into ignoring its safety guidelines). Existing defenses either reduce the model's usefulness or slow it down significantly. SafeSteer is a proposed defense method that uses singular value decomposition (SVD, a mathematical technique for breaking down data into simpler components) at inference time (when the model is running, without changing its internal weights) to identify and remove harmful instructions while keeping the model's normal abilities intact, achieving over 60% reduction in attack success rates with minimal slowdown.

Solution / Mitigation

SafeSteer uses singular value decomposition (SVD) to purify a low-dimensional "safety subspace" from noisy activation differences, then projects the raw steering vector into this subspace to isolate the core safety signal from noise and adaptively remove harmful influences while preserving the model's ability to handle benign inputs. The method is applied at inference time without modifying model weights and avoids iterative response generation.