零样本LLM护栏新思路：用冻结表示+拓扑推理构建极简安全流形，参数少还能防越狱。
arXiv:2608.08485v1 Announce Type: new Abstract: Current LLM safety guardrails face a fundamental tension: fine-tuning distorts pre-trained representat…
AI安全护栏正在阻碍进攻性网络安全研究人员获取零日漏洞，引发行业对随意决策的担忧。
We spoke with several cybersecurity researchers, who look for unknown vulnerabilities and develop tools to exploit them, about how OpenAI’s and Anthro…
Guard Vector通过任务向量组合与流式感知前缀SFT，突破英语限制实现多语言LLM安全护栏，提供可扩展的防护新范式。
arXiv:2509.23381v2 Announce Type: replace Abstract: We introduce Guard Vector, a safety task vector computed as the parameter difference between a gua…
ECCV 2026新作：输出感知安全护栏，精准解决多模态大模型“宁可错杀一千”的过度拒绝问题。
arXiv:2607.09697v1 Announce Type: new Abstract: Existing safety mechanisms for multimodal large language models (MLLMs) face a fundamental trade-off b…
无需额外推理即可强化大模型安全，意图驱动的新型护栏训练方法。
arXiv:2607.06326v1 Announce Type: new Abstract: Large language models deployed in open-world applications require safety guardrails that are both robu…
AI代理直面生产数据库：架构设计、安全护栏与踩坑经验一次讲透
So we have arrived at a point, our agent neeeds to interact with real production database. So if you have implemented agents beyond the prototype stag…
用Zig打造的高性能命令卫士，本地守护编码代理，86个安全包拦截危险操作，MCP调用本地分类，会话不离开电脑。
Article URL: https://github.com/christopherkarani/Orca Comments URL: https://news.ycombinator.com/item?id=48782806 Points: 2 # Comments: 0
基于神经元激活的隐私护栏，让LLM代理安全守护用户数据，不泄露敏感信息。
arXiv:2601.14660v2 Announce Type: replace-cross Abstract: Agentic Large Language Models (LLMs) are models able to reason, plan, and execute tools over…
AI浏览器新漏洞被曝光，通过“催眠”式提示绕过安全护栏，引发对AI风险的重磅警示
Telling an LLM that 2 + 2 = 5 is enough to make it follow forbidden instructions.
AI不是黑客的万能工具，安全团队正借它扭转攻防态势——破解护栏难题后的新视角
Article URL: https://badshah.io/blog/ai-is-the-best-thing-to-happen-to-security/ Comments URL: https://news.ycombinator.com/item?id=48705198 Points: 1…
为印度语言打造多语言安全护栏，填补非英语场景内容安全空白。
arXiv:2606.22841v1 Announce Type: cross Abstract: As Large Language Models (LLMs) achieve widespread integration across diverse linguistic landscapes,…
多模态大模型安全新方案：提出动态推理的政策自适应护栏，应对VLMs在消费、医疗等场景的部署风险。
arXiv:2606.22873v1 Announce Type: cross Abstract: Vision-language models (VLMs) are increasingly deployed in consumer, medical, financial, and enterpr…
面向中文大模型安全，提出细粒度护栏CHILLGuard，可扩展数据构建配合模型感知偏好对齐，为LLM安全检测提供新方案。
arXiv:2606.15396v1 Announce Type: cross Abstract: Malicious content generated from large language models (LLMs) could pose severe safety risks and eth…
Anthropic明日发布Claude Fable 5，号称最先进公开模型，具备Mythos级能力与强化安全护栏。
Anthropic will be releasing Claude Fable 5 tomorrow. It will be most advanced publicly accessible model, with Mythos-class capabilities and enhanced g…
从风险分类到行动计划修复，提出基于护栏反馈驱动的LLM Agent安全框架，解决行动前的风险评估与防御。
arXiv:2606.05805v1 Announce Type: new Abstract: LLM-based guardrails typically safeguard agents by evaluating proposed actions or inputs before execut…
提出 ConsisGuard 方法，对齐安全审议与策略执行，提升大模型护栏的可靠性与一致性。
arXiv:2605.31073v1 Announce Type: new Abstract: Reasoning-based LLM guardrails improve safety moderation by generating explicit rationales before issu…
引入潜在推理机制实现更高效鲁棒的AI安全护栏，降低计算开销。
arXiv:2605.29068v1 Announce Type: new Abstract: Maintaining the safety of large language models (LLMs) is crucial as they are increasingly deployed in…
ICML 2026收录：为LLM智能体设计的分层记忆增强安全护栏，提升复杂场景下的行为可控性。
arXiv:2605.05704v2 Announce Type: replace-cross Abstract: Recent advances in foundation models have transformed LLMs from passive conversational syste…
最新研究揭示：测试时训练（TTT）能绕过AI安全护栏，引发对模型防御机制的新思考。
arXiv:2605.22984v1 Announce Type: cross Abstract: Test-Time Training (TTT) is an emerging paradigm that enables models to adapt their parameters durin…
揭秘多语言下AI安全护栏失效的根本原因，从数据分布到tokenization的深度剖析。
arXiv:2605.17173v1 Announce Type: cross Abstract: Large language models exhibit safety degradation in non-English languages. Standard evaluation relie…