The State of Ransomware Q2 2026
For the past year, the ransomware conversation has centered on concentration: a handful of dominant RaaS operations cont
2026-08-14 归档
今日共收录 1589 条安全情报，包含 784 个 CVE，406 篇安全通告，以及 115 篇研究论文。
← 返回归档列表Apple sends new ‘Threat Notification’ alerts over mercenary spyware attacks
You're not alone if you just received an "Apple Threat Notification" saying it detected a "mercenary spyware attack targ
Ukraine shuts down 94 fraudulent call centers, seize millions in cash
Authorities in Ukraine shut down 94 fraudulent call centers across the country that lured people into investment scams o
Akira hackers disable EDR with Safe Mode, steal data but fail to encrypt
An Akira ransomware affiliate disabled the endpoint detection and response (EDR) solution on a compromised system by res
Hackers breach govt webmail while running parallel crypto fraud
The Jewelbug hacker group has been carrying out espionage operations targeting governments and militaries while also eng
Microsoft patches LegacyHive Windows zero-day vulnerability
Microsoft has released security patches to address a Windows zero-day vulnerability known as "LegacyHive," disclosed aft
AI 'watermark removers' flood the web. Almost none can prove they work.
Multiple 'watermark removers' have surfaced days after Anthropic began watermarking text generated by Claude, including
Trezor discloses data breach affecting nearly 14,000 customers
Hardware wallet manufacturer Trezor disclosed a data breach affecting nearly 14,000 of its customers after ShipMonk, its
Who Vets AI’s Code? The Scale Challenge Facing Open Source Ingestion
AI coding tools can introduce unvetted or hallucinated open source dependencies faster than traditional security reviews
White House taps security firms for offensive hack-back operations
A new White House memo signed by U.S. President Donald Trump instructs the National Coordination Center (NCC) to establi
WhatsApp rolls out new feature that flags potential scam messages
WhatsApp has begun rolling out a new optional "Scam Alert" feature, which uses a local machine learning model to warn us
Stealing Reasoning Traces from Proprietary LLM APIs
本文揭示了大型语言模型（LLM）提供商在客户端保存加密推理轨迹所引入的严重安全漏洞。许多主流提供商（如 Anthropic、OpenAI、Google）为避免泄露模型内部推理过程（即思维链），通常将推理步骤以加密块形式返回给客户端，并要求客
SoK: From Generation to Consumption of Privacy Documents in Software Systems
本文是一篇关于隐私文档（如隐私政策、隐私标签等）的系统性综述（SoK），从软件工程视角出发，对隐私文档从生成到消费的全生命周期进行了统一审视。研究背景是：隐私文档作为数字服务披露数据实践和获取用户同意的核心机制，近年来研究范围已从传统隐私政
Non-Degenerate Risk Certification for Automated Security Decisions: A Decision-Contract Theory with ATT\&CK-Aligned Triage as a Worked Instance
本文提出了一种针对自动化安全决策的风险认证理论框架，核心动机是揭示并修复现有风险界在安全决策中的退化问题。作者首先证明，在传统无条件风险界的定义下，存在一种结构性缺陷：一个从不采取行动的决策选择器可以平凡地使风险界归零，这意味着形式上满足风
AI Guardrail Survival under Single-Cycle Agentic Self-Summarization
该论文研究长期运行的大语言模型（LLM）智能体在上下文压缩（context compaction）过程中，站立安全约束（standing safety constraints）如何丢失。智能体在长时间运行中会定期压缩上下文，用模型生成的摘要
An Empirical Study of Output-to-Input Loops for Black-Box Backdoor Detection in Fine-Tuned Open-Weight LLMs
本文提出并实证评估了一种名为 self-feeding（自馈送）的黑盒后门检测方法，用于检查从公共仓库下载的微调大语言模型（LLM）是否被植入后门。研究动机是：任何人都可以上传声称安全的微调模型，而用户往往没有训练数据、干净参考权重或触发短
RangeFactory: Scalable Construction of Multi-Hop Cyber Ranges
本文提出的 RangeFactory 是一个面向多跳网络靶场（cyber range）的自动化编排框架，核心目标是解决现有靶场构建无法自动将大量孤立漏洞环境组合为端到端验证的多跳攻击链的问题。现实中的网络攻击通常需要跨越多个主机和网段持续推
From Prompt Injection to Web Exploitation: Revisiting Classic Vulnerabilities in LLM-Integrated Applications
本文研究了大语言模型（LLM）集成应用中的一类新型安全威胁，即“LLM 中介的 Web 攻击”（LLM-mediated web attacks）。随着 LLM 越来越多地被集成到 Web 应用中，用于聊天机器人、工具调用和智能体流程，用户
Privacy-Preserving RAG by Concealing Sensitive Information from External LLMs
该论文针对检索增强生成（RAG）中一个常被忽视的隐私问题展开研究：外部生成模型（如第三方LLM）在回答用户查询时，能够同时访问用户的查询语句和检索到的文档，这些内容可能包含敏感信息，存在被滥用或用于非预期目的的风险。已有的RAG隐私研究主要
The Next Challenge for Agentic Cybersecurity: A Realistic, Contamination-Free Reverse Engineering Benchmark
本文针对人工智能智能体在网络安全能力评测中的一个关键盲区：现有评测大多基于源代码，而真实世界中大量高价值目标（如恶意软件、固件、专有应用）仅以二进制形式分发，必须经过逆向工程（RE）恢复程序语义后才能进行安全分析。然而，评估智能体逆向工程能