Claude Chrome extension flaw lets malicious extensions trigger AI actions
A flaw in Anthropic's Claude for Chrome browser extension could allow a malicious extension to trigger predefined AI act
2026-07-17 归档
今日共收录 1302 条安全情报，包含 372 个 CVE，507 篇安全通告，以及 120 篇研究论文。
← 返回归档列表New OkoBot framework deploys 20 payloads to steal data, crypto
A new malicious framework called OkoBot is delivering more than 20 payloads in attacks focused on stealing cryptocurrenc
23andMe to pay $18 million in new genetics data breach settlement
Genetic testing company 23andMe has agreed to pay $18 million to settle claims from a coalition of 43 attorneys general
Scattered Spider members behind TfL hack get five years in prison
Two leading members of the Scattered Spider cybercrime collective were sentenced to five years and six months in prison
Windows 11 24H2 Home and Pro reach end of support in 90 days
Microsoft announced on Wednesday that systems running Windows 10 Enterprise LTSB 2016 and Home and Pro editions of Windo
CISA orders feds to patch actively exploited Oracle flaw by Saturday
CISA has ordered federal agencies to secure their systems by Saturday against ongoing attacks exploiting a critical vuln
Russian hackers trojanize WebEx, Zoom apps to push Starland malware
A financially motivated Russian threat actor tracked as UAT-11795 is using trojanized software to steal credentials and
New Spirals ransomware encrypts victim network in under 24 hours
A new ransomware actor called Spirals completed a corporate intrusion, from initial access to data theft and encryption,
ThreatsDay: Game Cheat Spyware, 24-Hour Ransomware, Chrome Sync Stalking + 12 More Stories
A lot of this week’s trouble starts with something that looks close enough. A familiar repo. A useful installer. A harm
Oak Emerges From Stealth Mode With $60 Million in Funding
The startup has built an AI-powered Identity Operating System that governs all identities across an organization’s envir
When Binaries Talk Back: Representation-Confusion Attacks on LLM-Assisted Reverse Engineering
该论文研究了LLM辅助逆向工程（RE）流水线中的一种新型安全威胁：表示混淆攻击（Representation-Confusion Attacks, RARE）。攻击者可以通过构造恶意二进制文件，使得RE系统错误地将二进制中提取的数据视为指令
MemPoison: Uncovering Persistent Memory Threats and Structural Blind Spots in LLM Agents
该论文研究了LLM智能体中持久化外部记忆机制带来的安全威胁。智能体通过持久记忆实现跨轮次连续性，但这也引入了安全漏洞：攻击者可通过标准交互渠道注入恶意内容，这些内容被保留并在后续轮次中扭曲智能体行为。为此，作者提出了MemPoison，一个
Is External Database Protection Static in Retrieval-Augmented Generation? Rethinking Privacy Preservation under Dynamic Queries
该论文针对检索增强生成（RAG）系统中隐私保护方法的静态风险假设问题，提出了一种动态分层差分隐私框架（PA-HDP）。现有隐私保护方法通常假设所有检索文档具有相同的隐私泄露风险，但RAG系统中文档的隐私风险高度依赖用户查询，即隐私泄露具有查
When Words Are Safe But Actions Kill: Probing Physical Danger Beyond Text Safety in Hidden-State Risk Space
大型语言模型（LLM）越来越多地被用作具身智能体的高级规划器，在此场景中，语言上看似安全的指令在物理世界中可能变得危险。本研究探讨了这种物理层面的危险是否与传统的文本级内容危险属于同一安全问题。通过隐藏状态方向分析和随机分裂空测试，作者发现
DataShield: Uncovering Risky Fine-Tuning Data Across LLMs Through Consensus Subspace Alignment
本论文针对微调大语言模型（LLMs）时，即使使用良性任务特定数据也可能显著削弱模型安全性的问题，提出了一种名为DataShield的数据评估框架。现有方法通常依赖单个模型及其分词器计算的均值向量来表征安全方向，导致风险评估的有效性和可迁移性
Fully Automated End-to-End Adversary Emulation from MITRE ATT\&CK Based Cyber Threat Intelligence Using LLMs
本文提出了一种完全自动化的端到端对手模拟框架，利用大型语言模型（LLM）从符合MITRE ATT&CK标准的网络威胁情报（CTI）报告生成、执行并恢复攻击剧本。与先前工作（如AURORA）不同，该框架无需人工干预，统一了剧本生成、执行和故障
Context Contamination in LLM Analysis of Network Security Logs: Poison with Passive Prompt Injection and Mitigation Evaluation
本文研究了大型语言模型（LLM）在安全运营中心（SOC）日志分析任务中的安全漏洞。随着LLM被广泛用于日志摘要、告警分类和威胁调查，它们需要处理来自外部服务的日志数据，并将其作为自然语言上下文进行分析。论文指出，这种架构引入了一种关键脆弱性
Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Values
该论文揭示了大型语言模型（LLMs）中存在的一种名为“价值泄露”（Value Leakage）的现象：模型在回答用户问题时，其自身隐含的价值观会无意识地影响生成的内容，且这种影响通常不会向用户披露。例如，当用户询问AI泡沫破裂的概率时，Cl
Agent Skill Security: Threat Models, Attacks, Defenses, and Evaluation
随着大型语言模型（LLM）智能体的广泛应用，可复用的技能模块（skill）成为构建智能体能力的基础单元，这些技能能够被封装、共享并在不同应用间复用。然而，现有的安全研究主要集中在提示注入和运行时执行上，对技能全生命周期中的安全风险探索不足。
Traffic-Aware Randomized Smoothing for LLM-Based Network Intrusion Detection
本文针对基于大语言模型（LLM）的入侵检测系统（IDS）在实际部署中面临流量操控攻击的鲁棒性问题，提出了一种流量感知随机平滑（TA-RS）认证防御方法。现有研究大多通过实验验证了LLM-IDS对恶意扰动的脆弱性，但缺乏理论保证。TA-RS是