Microsoft enables Windows settings backup by default for orgs
Microsoft announced that Windows settings backup and restore is now enabled by default on all Microsoft Entra-joined or
2026-10-02 归档
今日共收录 1399 条安全情报，包含 504 个 CVE，518 篇安全通告，以及 100 篇研究论文。
← 返回归档列表Hackers stole Pentagon personnel records of over 3 million people
The Pentagon's Defense Manpower Data Center (DMDC) is notifying millions of military service members that hackers stole
Metamask discloses security incident affecting its infrastructure
On Thursday, cryptocurrency wallet provider MetaMask has disclosed an ongoing infrastructure security incident affecting
Russian state hackers use new RedFlick technique to push malware
The Russian state actor Star Blizzard has been using a new malware installation tactic dubbed "RedFlick" to deploy its s
DIVD says Zammad zero-days enabled AI-driven network breach
The Dutch Institute for Vulnerability Disclosure (DIVD) says that the breach of its network was possible by exploiting a
Police Arrest 16-Year-Old Suspected of Running KillSec, Seize Ransomware Leak Site and Servers
Police in Spain have arrested a 16-year-old whom investigators suspect of running the KillSec ransomware group. KillSec
ThreatsDay: AI-Powered Zero-Day Chain, 543K Live Secrets, Model Inspection RCE and 13 More Stories
This week, the useful words are boring ones: inspect, cache, compile, store, trust. Each sounds harmless. Each can becom
How Financial Services Companies Can Modernize Their Software Supply Chain
Every security leader at a bank, insurer, or asset manager has had a version of this conversation: Security wants to eli
OpenAI Disrupts Reasoning Extraction Campaign Linked to Moonshot AI Associates
OpenAI on Wednesday said it identified and disrupted a coordinated distillation campaign that was designed to illicitly
Police Shut Down KillSec Ransomware, Identify Alleged Teen Leader
Police took control of KillSec’s leak site and secured at least 110 terabytes of data stolen from victims. The post Poli
Treasury Blacklists Most-Wanted ATM Malware Developer and His Network
The US government continues its crackdown on Tren de Aragua over its ATM jackpotting scheme. The post Treasury Blacklist
PACE: Provenance-Aware Capability Enforcement for Tool-Using LLM Agents
该论文聚焦工具调用型大语言模型（LLM）智能体的运行时安全问题。此类智能体把模型生成的文本转化为真实世界的副作用（文件写入、API 调用、消息发送等），因此被投毒的工具元数据、检索到的网页、记忆内容以及可复用的技能（skill）都可能操纵下
Representation Transitions Reveal Emerging Safety Risks in Multi-Turn LLM Agents
本文研究多轮（multi-turn）攻击对智能体（agentic）系统的威胁以及如何在运行时进行防御。作者指出，多轮攻击的关键特征是「组合性」：攻击者可以把每一步单独看都合乎许可的动作，编排成最终有害的结果，这使得那些只孤立评估单个动作或单
ActionGuard: Tool Call Authorization under Poisoned Skills
基于大语言模型的智能体通过第三方技能(skill)扩展能力,这些技能通常包含任务指令、脚本和工具调用流程。风险在于:即使一个技能本身看似良性,只要其中被植入少量恶意指令,一次完全正常的用户请求就可能被诱导成危险的工具调用,例如数据外泄、删除
Can Agents Trust Their Skills? Uncovering Unsafe Chains of Trust in Skill-Based LLM Agents
该论文研究了基于「技能(Skill)」的可安装能力包给 LLM 智能体带来的信任链安全问题。当前 LLM 智能体生态中,技能以指令、代码与资源打包的形式分发,用户安装后,智能体框架会在后续任务中自动调用这些技能。这形成了三层信任传递:用户把
Pretext: Defeating Malicious Skill Detection Frameworks for AI Agents
该论文研究 AI Agent(如 OpenClaw、Claude Code)技能(Skill)生态中的恶意技能检测问题。技能通过向 Agent 上下文注入指令与信息来扩展其能力,已成为 Agent 能力扩展的主流方式;已有研究表明第三方技能
Autonomous OSS Threat Detection via Taxonomy-Aligned LLMs
开源软件（OSS）生态系统面临日益复杂的供应链攻击，包括域名仿冒（typosquatting）、依赖混淆、Trojan Source 混淆、恶意构建注入以及 CI/CD 管道投毒。现有检测方法依赖签名和规则系统，难以泛化到攻击变种和新兴威胁
The Innocent Courier: Covert Exfiltration Through Legitimate LLM Web Fetching
该论文关注 LLM 与 LLM Agent 在真实工作流中的安全与隐私风险。此前研究多聚焦于提示注入、以及用户数据被聊天机器人服务商获取等问题，并提出了输入结构化、本地部署模型等缓解手段。但作者指出一个被忽视的方向：机密数据可能经由 LLM
Safety Must Survive Self-Improvement: Why Failures Persist and How Agents Recover
该论文研究递归自我改进（RSI）场景下智能体安全性的维持问题。所谓 RSI，是指智能体可以把自身的有用改动跨"代"传递下去，从而不断演化。作者指出，让安全在这种跨代演化中存活，包含两个必须同时解决的挑战：一是防止不安全行为在代际之间持续存留
Identity-Bound Governance Under Execution Uncertainty: An Accountability Proof Block for LLM Agent Persistent Halts, with Cryptographic Implementation and Cross-Model Calibration
该论文研究 LLM 智能体在自动化治理中的一类特殊失效状态：智能体同时拒绝了「继续执行」与「自动停机」两种选项。作者将这一状态命名为「身份绑定的治理事件」（identity-bound governance event），其触发条件是系统已