FBI tells ShinyHunters members to turn themselves in after recent arrest
The FBI is warning members of the ShinyHunters extortion group to turn themselves in after Dutch police arrested a man t
2026-09-30 归档
今日共收录 1598 条安全情报，包含 728 个 CVE，467 篇安全通告，以及 100 篇研究论文。
← 返回归档列表Former US Air Force members sent to prison over BEC attacks
Two former members of the United States Air Force were sentenced to a combined 189 months in federal prison for their ro
Windows 11 2026 Update released, here's everything you need to know
Microsoft has started rolling out Windows 11 26H2 to everyone, and while it's this year's big annual feature update, you
New Spectre v2 attack variant leaks Linux root password hash in minutes
A new Branch Target Reuse (BTR) attack has been devised that can recover root password hashes on Intel computers running
Automated AI agent used to breach cybersecurity nonprofit DIVD
The Dutch Institute for Vulnerability Disclosure (DIVD) suffered an AI-driven cyberattack that the organization describe
Vietnamese man charged in $16 million 'pig butchering' crypto scam
A Vietnamese national was charged with money laundering for his role in a massive "pig butchering" scam, which defrauded
Kiteworks patches critical flaw, brings customer systems online
American tech company Kiteworks has lifted a precautionary advisory asking customers to shut down systems after patching
Apple patches CoreGraphics zero-day flaw exploited in attacks
Apple released security updates to fix a zero-day vulnerability exploited in "extremely sophisticated" targeted attacks
Rig Security Emerges From Stealth With $12M to Tackle Agentic AI Identity Risks
Rig provides an identity dependencies graph to distinguish between legitimate users and rogue AI agents The post Rig Sec
CounterSteer: Suppressing Indirect Prompt Injection with Activation Steering
该论文针对 LLM 智能体面临的间接提示注入（indirect prompt injection）问题——即被检索回来的不可信文本被模型当作指令执行，从而导致工具调用被劫持——提出了一种推理时（inference-time）的模型内部防御方
Same Bytes, Different Authority: Reserved-Token Representations in Chat-Template Prompt Injection
该论文研究 LLM 智能体场景下的提示注入（prompt injection）问题，聚焦一个以往被忽视的细节：当注入指令被模型自身的聊天模板标记（如 <|im_start|> 这类保留控制标记）包裹时，攻击效果会显著增强。核心观察是，同一段
pikit: A Composable Toolkit for Indirect Prompt Injection Research and Evaluation
该论文提出 pikit，一个面向间接提示注入（Indirect Prompt Injection, IPI）的可组合研究工具包，用于系统化评测 LLM 智能体面临的此类威胁。间接提示注入的核心问题是：攻击者将恶意指令嵌入智能体从外部检索到的
Render Before Reading: Visual Rendering as a Prompt Injection Defense
该论文研究大语言模型（LLM）面临的提示注入攻击，重点考察对抗内容的输入模态对攻击效果的影响，并提出一种免训练的防御方法。作者首先提出一个关键观察：多模态 LLM 在对抗性指令以纯文本形式出现时，比同一指令通过非文本通道（例如渲染成图像，或
"Nothing to See Here'': Unintended Disclosure through Revision Traces of LLM Deliverables
该论文研究大语言模型助手在协助用户起草发给第三方的交付物时出现的一类隐私泄露问题。用户在私密草稿阶段可能先写入某个条目、随后又删除或替换它；模型虽然在目标内容中删掉了该条目，却在陈述本次编辑动作时把它重新说了出来，作者将这类语句命名为"修订
Where Do LLMs Decide to Break the Rules? Mechanistic Localization of Prompt Injection Compliance
该论文研究的是提示注入（prompt injection）攻击成功时，大语言模型内部究竟在何处「做出违反系统角色设定的决定」这一机制性可解释性问题。作者指出，既有工作大多停留在量化攻击成功率（即模型有多频繁地服从对抗性指令），而忽视了决策发
SEABench: Benchmarking Endogenous Misalignment In Self-Evolving Agents
该论文关注自演化 LLM 智能体（self-evolving LLM agents）带来的内生性失配（endogenous misalignment）风险。这类智能体在部署后能够根据用户与环境反馈修改自身"装备"（harness），包括控制
PrivacySkills: How Privacy Guidance Shapes Source Selection in LLM Agents
本文提出 PrivacySkills，一个受控评测框架，研究大语言模型智能体在信息获取路径选择上如何受隐私指引呈现方式影响。核心问题是：当多种获取路径能带来同等任务相关价值时——查询公开可得的个人信息、访问机密来源，或直接与用户交互——智能
When Valid Tool Calls Change Meaning: Formation-Consistent Dispatch for LLM Agents
该论文聚焦于工具增强型 LLM 智能体(tool-enabled agents)在调用分派环节的一类结构性安全缺陷。作者指出：智能体在生成工具调用时，依据的是模型可见的接口描述(descriptor，即 schema、名称、参数说明)，而宿
Making Duplicate Reimbursement Unrepresentable: A Verified Ethereum E-Invoice System for Humans and AI Agents
该论文针对电子发票在消费/报销侧长期存在的三个痛点——同一张发票可被重复提交报销、收票方难以验证真伪、数据集中托管在中心化机构从而形成性能瓶颈与单点故障——提出并完整实现了一套基于以太坊的电子发票系统。方法上，作者首先将发票生命周期形式化为
ToolFence: Fine-Grained Authorization for Secure Tool-Using LLM Agents
该论文关注工具调用型 LLM Agent 面临的一个根因性问题：在单一上下文窗口中，来自用户的「可信指令」与来自外部环境的「不可信观测」（网页、邮件、工具返回内容等）混杂在一起，因此间接提示注入可以借观测内容劫持 Agent 的推理，甚至反