This month in security with Tony Anscombe – August 2026 edition
Details about the Hugging Face hack, critical infrastructure under attack, a spoofed in-flight Wi-Fi network, and more o
2026-09-01 归档
今日共收录 1266 条安全情报，包含 424 个 CVE，451 篇安全通告，以及 107 篇研究论文。
← 返回归档列表Cronos blockchain restarts after $74 million Tectonic exploit
The Cronos blockchain network has resumed trading activity after a price-manipulation attack on the Tectonic cryptocurre
OpenAI confirms ChatGPT outage as users report errors
ChatGPT Work is experiencing a partial outage, and users across multiple subscription plans may be unable to start or co
File servers are here to stay. Here’s how to manage them securely
File servers remain a critical part of many IT environments, but managing access securely can become complex as permissi
Berlin confirms data theft after Rhysida ransomware attack claims
Berlin's city administration has confirmed that cybercriminals are attempting to extort the city after the Rhysida ranso
Microsoft says Windows 11 KB5120998 update resets mouse settings
Microsoft has confirmed that mouse settings are being reverted on Windows 11 systems after installing the KB5120998 Augu
Nigerians extradited to US for sextortion, deaths of two teens
Two Nigerian men extradited to the U.S. on Thursday have been charged with involvement in sextortion schemes that result
Microsoft asks users to ignore 'Antivirus is turned off' errors
Microsoft asked customers this week to ignore alerts that Defender Antivirus has been turned off after installing the la
Securing Claude Code: The New Compliance API, Local Visibility, and Identity Governance
Claude Code reads files, runs shell commands, invokes MCP tools, and acts through the credentials available on a develop
McKesson Confirms Data Breach as Attacker Deadline Looms
The ShinyHunters extortion group has claimed the theft of 284 million records from the company’s systems. The post McKes
Judge Says Pentagon’s Measures Against Anthropic Were ‘Illegal and Baseless’
The ruling is part of Anthropic's legal battle against the Pentagon after the government labeled the company as a supply
Recognition Without Enforcement: Configuration-Dependent Failures in LLM Agent Instruction Arbitration and External Control
该研究针对 LLM 代理在指令仲裁中的安全边界问题。LLM 代理需要同时处理系统提示、用户输入、记忆和工具等来源的指令，但研究发现模型对指令来源的仲裁不能保证强制信任边界，存在一种'识别-执行差距'：模型激活中可以线性解码出源格式特征（如角
WoE Wrote It? Watermarking Mixture-of-Experts LLMs for Black-Box Text Provenance
大型语言模型（LLM）水印技术为文本溯源提供了机制，使模型所有者能够识别机器生成内容并归属到特定水印模型。然而，现有LLM水印方法主要依赖推理时采样器，且聚焦于密集模型。这类方法仅在文本通过模型所有者受控API显式生成时有效，在模型被窃取或
Layered LLM Defenses as an Ensemble: Access Tiers, Inference Cost, and the Measured Failure Correlation Between Defense Layers
本文研究了大语言模型（LLM）防御体系的分层堆叠效果。实践中，防御者常通过叠加多层防御来增强安全性，但堆叠是否真正有效取决于一个关键条件：各防御层必须在不同的失败样本上出错，即失败独立性。然而，LLM安全文献虽推荐这一假设，却从未实际测量。
Moirae: A Multimodal Agent Collaborative Framework for Dynamic Android Malware Detection
本文提出了一种名为 Moirae 的多模态智能体协作框架，用于动态 Android 恶意软件检测。针对现有机器学习检测器容易受到概念漂移影响的问题（因为依赖随时间变化的实现特定特征），以及当前基于大语言模型（LLM）的检测器通常仅依赖代码或
Breaking Darknet CAPTCHAs with general purpose LLM
本文评估了通用大语言模型（LLM）在破解暗网环境中常见CAPTCHA验证码任务上的有效性。暗网CAPTCHA通常设计为无JavaScript运行，因此与主流验证码系统存在显著差异。研究选取了三种具有代表性的挑战类型：开放圆圈定位、旋转对齐和
Identity by Design, Demographics by Accident: Demographic Leakage and Suppression in Behavioral Biometric Embeddings
本文首次系统性地审计了行为生物特征认证（BBA）系统中的人口统计信息泄漏问题。BBA 系统使用深度学习模型将眼动、语音、按键/触摸动态、步态等生物信号转换为身份嵌入向量，用于用户认证。然而，这些嵌入在编码身份的同时，可能无意中泄露性别、年龄
The Fragility of Jailbreak Robustness Across Operational States
现有越狱评测通常仅报告默认配置（香草状态）下的单一攻击成功率（ASR）。但实际使用中，用户与聊天模型的交互会引入多种操作状态，例如不同的系统提示、上下文或参数。本文发现，即便攻击方法固定，仅仅改变一条与安全无关的普通系统提示，也会显著改变A
SingProbe Technical Report
SingProbe 技术报告提出了一种面向大语言模型（LLM）部署的轻量级内在运行时护栏方法。传统的外部护栏模型通常独立于主模型运行，会引入额外推理成本、延迟安全信号，并且与越来越强大的基础模型存在能力不匹配的问题。SingProbe 直接
ECLIPSE: Self-Evolving Stealthy Prompt Injection Attack against Long-Horizon Agentic Systems
本文提出了一种针对长周期（long-horizon）LLM Agent 系统的自进化、隐蔽性提示注入攻击框架 ECLIPSE。随着 Codex、Claude Code、OpenClaw 等 LLM Agent 通过反复调用工具来规划和执行复