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

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

As the capabilities of Vision Language Models (VLMs) continue to improve, they are increasingly targeted by jailbreak attacks. Existing defense methods face two major limitations: (1) they struggle to ensure safety without compromising the model’s utility; and (2) many defense mechanisms significantly reduce the model’s generation efficiency. To address these challenges, we propose SafeSteer, a lightweight inference-time steering framework that effectively defends against diverse jailbreak attacks without modifying model weights. At the core of SafeSteer is the innovative use of singular value decomposition (SVD) to purify a low-dimensional “safety subspace” from noisy activation differences. By projecting the raw steering vector into this subspace, SafeSteer isolates the core safety signal from noise, adaptively removing harmful influences while preserving the model’s ability to handle benign inputs. SafeSteer avoids iterative response generation and introduces only limited overhead compared with other single-pass activation-steering defenses. Extensive experiments show that SafeSteer reduces the attack success rate by over 60% while maintaining the model’s utility on benign tasks, without introducing significant inference latency. These results demonstrate that robust and practical jailbreak defense can be achieved through simple, efficient inference-time control.

AlcaTRAz - Anchored Tree-Rule Defense Against Jailbreaks

Large language models (LLMs) are vulnerable to jailbreak attacks that bypass safety alignment through carefully crafted prompts. Many existing defenses require access to model weights or internals, making them difficult to apply to black-box deployments. We propose AlcaTRAz (Anchored Tree-Rule defense Against jailbreaks), a prompt-level defense based on rule trees that operates exclusively on the input text and requires no modification or retraining of the target model. The method automatically learns a transferable transformation rule that inserts controlled character-level perturbations at selected positions, thereby disrupting structural regularities exploited by jailbreak attacks while largely preserving the model's utility on benign queries. We evaluate the proposed method across 33 open-weight models, 22 jailbreak attack types, and a benchmark of short, single-turn benign questions, comparing against three representative prompt-level baselines (Llama Guard, RA-LLM, Goal Prioritization). Among the compared defenses, AlcaTRAz achieves the best composite security and functionality score in 73.4 % of model-attack combinations and shifts the aggregate score from a modal value of 10 (maximal-severity response to the malicious request) in the undefended setting to a modal value of 2 (near-refusal) after defense, while keeping the mean benign score within 0.27 points of the undefended baseline (8.35 vs. 8.62 on a 0-10 scale). AlcaTRAz substantially reduces but does not eliminate jailbreak success: a high-severity tail remains, and we do not consider adaptive attackers, so we position it as one layer within a defense-in-depth strategy rather than a standalone guarantee.

Fully Unleashing the Multimodal Attacker: Meta-Adaptive Jailbreaking of Vision-Language Models

The safety of large vision-language models is increasingly stress-tested by multimodal jailbreaks, yet existing attacks remain largely static at the meta level: template-based attacks freeze the image-text layout, while iterative attacks adapt only the image-text content with fixed attack strategies and frozen attacker parameters. We propose Meta-Adaptive Multimodal Jailbreaking (MAMJ), which instead optimizes the attacker itself along two axes: an attack strategy prompt (ASP) governing attack iteration and attacker model weights determining attack effectiveness. Across groups of multimodal attack trajectories, an LLM-based critique first refines the ASP, after which group-aggregated attack success rate (ASR) rewards update those weights. On MM-SafetyBench, MAMJ achieves 81.0%, 78.9%, and 82.3% ASR against GPT-4o, Gemini-3-Pro-Preview, and Seed 2.0, respectively, outperforming the strongest sample-level baseline by up to 24.1 percentage points. The learned attacker, comprising the optimized ASP and attacker weights, also transfers without retraining to unseen victims and remains effective under representative defenses. These results reveal a systemic vulnerability of frontier VLMs to meta-adaptive jailbreaks and motivate defenses against meta-level adversaries. Code is available at https://github.com/Alibaba-VELLDEPTH/MetaJailbreak-VLM.

NeuronGuard: Robust LLM Safety Alignment via Ablation-Aware Safety Signal Redistribution

A fine-tuning-stage defense that simultaneously hardens LLMs against both attack classes by redistributing safety signals across a broader set of neurons, and provides a formal guarantee that NeuronGuard strictly reduces the attack success rate (ASR) upper bound.

Jailbreaking Jailbreaks: A Proactive Defense for LLMs

SafeRI: Recognition and Intervention for Token-Level Safety Intervention in Large Vision Language Models

Existing safety alignment methods for vision-language models usually modify the model behavior globally: once the safety parameters are trained or loaded, they participate in both unsafe and already-safe generations. This always-on intervention can unnecessarily perturb the model's original reasoning path and degrade general multimodal capabilities. We argue that safety alignment should be an on-demand intervention rather than a permanent modification to every decoding trajectory. To this end, we propose a streaming recognition and gated LoRA framework for intrinsic VLM safety. During autoregressive generation, a lightweight recognizer estimates whether the current pre-token generation state is safe or unsafe. Its output updates the LoRA gate for the following decoding step; otherwise, generation follows the frozen-backbone policy. The LoRA module is trained from unsafe prefixes, transition statements, and safe continuations, so that it learns to redirect unsafe generations back to safe responses after activation. Experiments across multiple safety and general-purpose benchmarks demonstrate the effectiveness of our method in post-alignment settings.