Why Account Takeovers Are Rising and How to Stop Them
Account takeovers are rising as attackers bypass traditional defenses through phishing, session hijacking, and MFA fatig
2026-06-18 归档
今日共收录 819 条安全情报，包含 0 个 CVE，417 篇安全通告，以及 115 篇研究论文。
← 返回归档列表India's Telegram ban hit the UAE too. Here's how to get around it
India has banned Telegram until June 22 after the app was used to circulate leaked exam papers. CEO Pavel Durov accuses
Microsoft confirms Office apps launch issues after June updates
Microsoft is investigating a new issue preventing third-party applications from launching Microsoft Office applications
CISA orders feds to patch max severity Joomla plugin flaw by Friday
The U.S. Cybersecurity and Infrastructure Security Agency (CISA) has ordered federal agencies to patch a maximum-severit
Microsoft working on Defender patch for RoguePlanet zero-day
Microsoft confirmed that it's working on a security patch for a Defender zero-day vulnerability named "RoguePlanet," dis
Kodak confirms data breach claimed by ShinyHunters extortion gang
Kodak has confirmed that it's working with external cybersecurity experts to investigate a security breach after hackers
Malicious JetBrains Plugins Steal AI API Keys as Chrome Extensions Capture Chatbot Chats
Cybersecurity researchers have flagged a "coordinated malware campaign" on the JetBrains Marketplace that has published
Microsoft Teams Relay Servers Abused in DragonForce Ransomware Attack
The attackers deployed a new Go-based backdoor that uses Microsoft Teams servers for command-and-control. The post Micro
MIDS: Detecting Stealthy Masquerade and Tampering Attacks on CAN Bus via Bidirectional Mamba
现代汽车中的电子控制单元（ECU）通过控制器局域网（CAN）总线进行通信，但CAN协议缺乏加密和认证，容易受到多种安全威胁。现有的入侵检测系统主要针对通过帧注入实现的制造攻击（如拒绝服务、模糊测试、标识符欺骗），这类攻击可以通过基于每个标识
Ghost Vectors: Soft-Deleted Embeddings Remain Reconstructible in HNSW Vector Databases
该论文研究了基于 HNSW（Hierarchical Navigable Small World）图结构的向量数据库中软删除操作导致的安全漏洞。在检索增强生成（RAG）流水线中，LLM 需要访问外部私有语料库，而 HNSW 数据库被广泛用于
Invisible Manipulation Channels in AI-Assisted Financial Advisory: Implications for Market Integrity and Regulatory Design
该论文研究AI辅助金融咨询（如信用评估、投资顾问）中存在的隐蔽操纵通道。作者发现LLM推理的采样层存在一个漏洞，允许攻击者在保持与输出审计机制（包括统计水印）完全兼容的情况下，系统性偏置AI生成的金融意见。这种推理阶段的操纵在统计上难以检测
Understanding and Mitigating Prompt Leaking Attacks in Real-World LLM-Based Applications
该论文针对基于大型语言模型（LLM）的实际应用中的提示泄露攻击进行了系统性研究。首先，作者指出系统提示（system prompts）是LLM应用的核心逻辑和开发者约束的载体，属于重要的知识产权，但容易受到提示泄露攻击。虽然已有研究在受控环
Evaluating the Effectiveness of LLMs in Aiding Compliance Testing of PKCS#1-v1.5
该论文研究了将基于语法的变异与大型语言模型（LLM）代码合成相结合的方法，以用于二进制协议的合规性测试。作者聚焦于PKCS#1 v1.5签名验证标准——一种广泛部署的基于TLV（类型-长度-值）编码的协议，并利用形式化验证的测试预言（Mor
CodeSentinel: A Three-Layer Defense Against Indirect Prompt Injection in Code Contexts
本文聚焦于代码大语言模型（Code LLM）面临的间接提示注入（Indirect Prompt Injection）威胁。当Code LLM从代码仓库、文档、问题线程或编程代理环境中检索外部代码上下文时，攻击者可以在注释、字符串、标识符或诱
PhantomSkill: Malicious Code Injection in Agent Skill Ecosystems
本文提出了一种名为 PhantomSkill 的新型攻击框架，针对基于大型语言模型（LLM）的编码代理技能生态系统。在当前的代理技能市场中，第三方开发者可以发布技能包来赋予 LLM 代理特定领域的能力，但这种机制也引入了供应链安全风险。传统
OpenAnt: LLM-Powered Vulnerability Discovery Through Code Decomposition, Adversarial Verification, and Dynamic Testing
自动化漏洞发现在大规模代码库中仍然充满挑战：传统静态分析误报率高，而模糊测试等动态方法需要大量基础设施且通常只针对特定类型的漏洞。近年来，大型语言模型（LLM）在程序行为的语义推理方面展现出潜力，但将LLM应用于仓库级安全分析会带来上下文管
TGCM: Topic-Guided Generative Disentanglement of Interleaved APT Technique Sequences
在企业环境中，多个高级持续性威胁（APT）攻击活动经常同时发生，导致审计日志中来自不同攻击者的技术序列随时间交织在一起。这种设置自然地引出了未知数量交织序列分离（UKISD）问题：从交织的技术序列中恢复多个潜在攻击活动，同时推断其数量和技术
Code-Augur: Agentic Vulnerability Detection via Specification Inference
本文提出一种名为 Code-Augur 的新型智能体漏洞检测方法，核心思想是“安全规范优先”。该方法首先让智能体分析代码库中各组件，当认为某组件安全时，将推理过程中使用的局部不变量作为安全规范（assertion）嵌入源代码。然后，利用指导
Signature filtering: a lightweight enhancement for statistical watermark detection in large language models
该论文提出了一种名为“签名过滤”（signature filtering）的轻量级检测增强模块，用于提升大型语言模型（LLM）输出的统计水印检测性能。现有的水印检测器在信号较弱、文本重复或水印被编辑时常常失效。签名过滤模块不修改水印嵌入和文
CloakLM: Obfuscating GPU Memory Layout to Mitigate Model Ex-filtration for Serving
本文提出 CloakLM，一种软件层面的 GPU 内存混淆框架，旨在缓解大模型在第三方共享加速器基础设施上部署时的模型权重泄露风险。现有攻击如 Hermes 可通过被动监听 PCIe 流量无损重建 DNN 模型，TunnelS 则利用驱动级