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

Summary

SafeSteer is a lightweight inference-time steering framework that defends Vision Language Models against jailbreak attacks without modifying model weights. It uses singular value decomposition to purify a low-dimensional safety subspace from noisy activation differences, then projects the raw steering vector into that subspace. The authors report a reduction of over 60% in attack success rate while maintaining utility on benign tasks.

Topics

• Prompt injection and jailbreaks

Related items

• InfoGem5-Shark: Smaug-based Heterogeneous Deep Learning Accelerator Security and Robustness Evaluation and Characterization FrameworkSimilar attack · OpenAlex (peer-reviewed AI security)
• MediumCVE-2026-123851: JetBrains Junie GitHub Action prompt injection bypass via HTML encodingSimilar attack · NVD/CVE Database
• InfoGuardrailAgent: An Agentic Framework for Jailbreak DefenseSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• MediumCPython flaw exposes AI inference servers to sandbox escape attacksSimilar attack · CSO Online
• InfoJev as an LLM Safety Guardrail: An Empirical Study of Input Filtering, Selective Output Review, and Over-RefusalSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)