Microsoft Self-Service Password Reset abused in Azure data theft attacks
A threat actor targeting Microsoft 365 and Azure production environments is stealing data in attacks that abuse legitima
2026-05-20 归档
今日共收录 1110 条安全情报，包含 284 个 CVE，411 篇安全通告，以及 113 篇研究论文。
← 返回归档列表Microsoft plans to improve Windows 11 driver quality in 2026
Microsoft plans to raise the quality bar of Windows 11 drivers, as drivers "sit at the heart of every Windows experience
Microsoft blames macOS update for undismissible Teams location prompts
Microsoft has confirmed user reports that the Teams team collaboration app is displaying non-dismissible location prompt
New Shai-Hulud malware wave compromises 600 npm packages
Threat actors earlier today published more than 600 malicious packages to the Node Package Manager (npm) index as part o
7-Eleven confirms data breach claimed by the ShinyHunters gang
Convenience store chain giant 7-Eleven confirmed that its systems were breached in a cyberattack claimed by the ShinyHun
Webinar: The hidden bottlenecks in network incident response
IT teams are increasingly overwhelmed by alerts from disconnected systems, forcing responders to manually coordinate inv
Microsoft confirms patching issues in restricted Windows networks
Microsoft says customers in restricted network environments may encounter Windows Update failures after installing the J
Popular GitHub Action Tags Redirected to Imposter Commit to Steal CI/CD Credentials
In yet another software supply chain attack, threat actors have compromised the popular GitHub Actions workflow, actions
Verizon DBIR 2026: Vulnerability Exploitation Overtakes Credential Theft as Top Breach Vector
Verizon’s 2026 DBIR finds vulnerability exploitation has overtaken credential abuse as the leading breach vector, as AI
Microsoft Disrupts Malware-Signing Service Run by ‘Fox Tempest’
Fox Tempest provides a service that cybercriminals use to distribute ransomware and other malware disguised as legitima
Cyber Resilience is the New Business Continuity Plan
The organizations best prepared to face disruption are those that align security, continuity and risk management around
Babel: Jailbreaking Safety Attention via Obfuscation Distribution Optimized Sampling
尽管大语言模型（LLM）经过严格的安全对齐训练，仍容易受到越狱攻击。现有黑盒方法多依赖启发式模板或穷举尝试，缺乏机制可解释性和查询效率。本研究揭示了LLM安全机制的内在脆弱性：安全对齐仅依赖于少量稀疏分布的注意力头，导致大部分表示空间监控薄
Trust No Tool: Evaluating and Defending LLM Agents under Untrusted Tool Feedback
本文研究了大型语言模型（LLM）代理在使用外部工具时面临的一种新型安全威胁：认知中毒（cognitive poisoning）。在这种攻击场景下，恶意工具在探索阶段表现得合理且有用，通过提供看似无害的反馈逐步积累代理的信任，只有当隐藏状态条
ASPI: Seeking Ambiguity Clarification Amplifies Prompt Injection Vulnerability in LLM Agents
本文研究了大型语言模型（LLM）代理（agent）在寻求澄清（clarification-seeking）行为下的安全性问题。澄清行为通常被视为代理的理想属性，允许其在执行不明确任务前先解决歧义。然而，作者发现这种交互模式会显著增加代理对提
Prompt2Fingerprint: Plug-and-Play LLM Fingerprinting via Text-to-Weight Generation
该论文针对大语言模型（LLM）广泛部署和再分发带来的模型来源追踪挑战，指出现有主动指纹注入方法（如微调嵌入身份信号）虽高精度且鲁棒，但存在严重的可扩展性瓶颈：每个新身份都需要独立的、资源密集的从头训练。为突破这一限制，作者提出首个将指纹注入
BiRD: A Bidirectional Ranking Defense Mechanism for Retrieval Augmented Generation
检索增强生成（RAG）在广泛部署的同时，也面临着对抗性投毒攻击的威胁。现有防御方法主要依赖语义分析或投票机制，存在计算成本高或鲁棒性不足的权衡，其根本局限在于仅关注语义内容相关性，而忽视了由排序结构定义的检索上下文。本文首先系统研究了毒化文
LLM Benchmark Datasets Should Be Contamination-Resistant
本研究聚焦于大型语言模型（LLM）基准数据集污染问题。基准数据集对于可重复、可靠且有区分度的LLM评估至关重要，但近期研究表明许多基准数据集已混入预训练语料（即被污染），这削弱了其作为模型泛化能力可靠指标的价值。论文首先广泛论证了基准数据污
Exploring and Developing a Pre-Model Safeguard with Draft Models
本文提出了一种基于小语言模型（SLM）的预模型安全防护设计，旨在高效检测针对大语言模型（LLM）的越狱攻击。研究首先系统性地分析了越狱攻击的可迁移性，特别是从LLM到SLM的迁移规律，发现关键影响因素。实验表明，针对LLM构造的越狱提示也容
Be Kind, Rewrite: Benign Projections via Rewriting Defend Against LLM Data Poisoning Attacks
大型语言模型（LLM）极易受到后门攻击（BA），攻击者通过向训练样本中注入触发器相关的有害内容来投毒数据。现有防御方法在广泛测试中效果不佳。本文探索了利用LLM重写作为主动防御手段来对抗数据投毒。首先，作者在理论上证明：当LLM重写利用开放
Prompts Don't Protect: Architectural Enforcement via MCP Proxy for LLM Tool Access Control
本文指出大型语言模型（LLM）在作为自主智能体时，能够从大型注册表中选择并调用工具，但存在关键安全漏洞：即使通过提示明确禁止，当未授权工具出现在智能体的上下文窗口中时，模型仍可能在对抗性场景下选择它们。作者提出了一种受控的MCP（Model