Tag
提出Counterfactual Neighborhood Risk Fields (CNRF)，一种基于证据的视觉-语言模型选择性安全防御框架，通过构建上下文匹配的反事实对来区分风险与陌生性，在三种VLM和五类多模态越狱攻击上实现86.06-89.30%的宏防御成功率，同时保持良性任务效用。
该论文提出一种训练无关的权重编辑方法，通过少量对比提示对直接修改模型权重，降低指令微调模型对有害指令（如说谎、性别歧视）的服从率，并在五个模型上验证了其优于激活引导等方法的鲁棒性。
Google DeepMind 发布了 Gemini 4 Argon，并详细介绍了围绕其网络安全与生物领域高能力所构建的安全防护体系，包括前沿安全框架、三层实时防护以及 Fairwind 计划的扩展。
This paper builds Guise, a cross-language benchmark showing that narrative wrapping dramatically increases LLM jailbreak success (up to 95.7% in Classical Chinese on Qwen3-1.7B), and proposes AXIS, a defense combining preference optimization with rotation and commitment objectives that achieves the best combined safety-usability score across Qwen3 and GLM-4 models.
HiRoute proposes a hierarchical routed prompt-tuning framework that separates category-agnostic safety control from category-specific response guidance for LLM safety alignment, reducing over-refusal while maintaining high safety rates.
Introduces Latent Personality Alignment (LPA), a lightweight adversarial training method that uses 66 psychometric personality statements to achieve near-zero attack success rates on jailbreak attacks without degrading utility, requiring only minutes on a single GPU.
CHASE introduces a co-evolutionary red-blue teaming framework that uses reinforcement learning to harden LLMs against adaptive black-box adversarial attacks, reducing jailbreak success by 43.2% on benchmarks while maintaining zero false refusals on benign prompts.
This paper introduces Anchored Bipolicy Self-Play, a method to improve AI safety by training distinct role-specific LoRA adapters on a frozen base model, addressing limitations in standard self-play red teaming.
This paper introduces a resource-efficient pruning framework that identifies and removes parameters associated with unsafe behaviors in large language models while preserving utility. Using gradient-free attribution and the Lottery Ticket Hypothesis perspective, the method achieves significant reductions in unsafe generations and improved robustness against jailbreak attacks with minimal performance loss.
ASGuard is a mechanistically-informed defense framework that mitigates jailbreaking attacks on LLMs by identifying vulnerable attention heads through circuit analysis and applying targeted activation scaling and fine-tuning to improve refusal behavior robustness while preserving model capabilities.
OpenAI proposes an instruction hierarchy approach to defend LLMs against prompt injection and jailbreak attacks by training models to prioritize system instructions over user inputs. The method significantly improves robustness without degrading standard capabilities.