Ch12 安全与治理¶

Agent 权限越大，安全责任越重：凭据、审计、合规

本章收录 130 篇实体，按深度递增排列。

本章导航¶

导读¶

AI Agent 正在获得越来越多的权限——执行代码、访问数据库、发送邮件、操作文件系统。

权限越大，攻击面越大。本章覆盖 Agent 安全的完整谱系：凭据管理（1Password 的机器身份方案）、Prompt 注入防御、供应链攻击（TanStack npm 事件）、恶意软件分析（GlassWASM WebAssembly 恶意代码）、逆向工程（Themida 脱壳）。

你还会看到 100 万+ AI 服务暴露在公网的扫描报告，以及 Google 与国际特赦组织联手打击商业间谍软件的行动。

安全不是"做完再考虑"的事——它应该内嵌在 Agent 架构的第一天。

本章内容¶

• 001. CISA urges critical infrastructure firms to 'fortify' before it's too late | Cybersecurity Dive
• 002. Token 撤销触发设备擦除的安全漏洞
• 003. A Framework for AI Threat Readiness
• 004. From SSH to REST: A Security-Driven Modernization of Slack's EMR Data Pipelines
• 005. Mozilla warns UK: Breaking VPNs will not magically fix Britain's age-check mess
• 006. Offensive Security Blog
• 007. Sandworm Hackers Shift From IT Breaches to Critical OT Targets
• 008. 5 Things to Know about the CLARITY Act
• 009. fedora hummingbird brings the container security model to a linux host os
• 010. 推出-amazon-bedrock-managed-knowledge-base助力企业人工智能应用程序更快速更准确
• 011. Mythos finds a curl vulnerability
• 012. 飞来汇借助 AWS Security Agent 构建跨境支付应用的智能安全防线
• 013. Canvas Hackers ShinyHunters Say Their Official Domain Was Suspended
• 014. Hermes Agent v0.14.0 核心架构与快速上手
• 015. Bleeding Llama：Ollama 未授权内存泄漏漏洞
• 016. SHub Reaper: macOS Stealer Spoofs Apple, Google, and Microsoft in a Single Attack Chain
• 017. LLMReaper - DOM Based AI Conversation Exfiltration via Browser Extensions
• 018. Resecurity | CVE-2026-20182: Unauthenticated Cisco SD-WAN Control-Plane Compromise via vHub Authentication Bypass
• 019. Towards Native Post-Quantum Private ETH - Privacy - Ethereum Research
• 020. Static Devirtualization of Themida
• 021. Static Devirtualization 2024
• 022. What My Privacy and Security Stack Actually Looks Like
• 023. Alliance for Critical Infrastructure (ACI): US Critical Infrastructure Cybersecurity Coalition
• 024. How an image could compromise your
• 025. Static Devirtualization of Themida
• 026. Inference Theft as AI Endpoint Attack Surface — Vercel Token Theft Defense 2026
• 027. Apple corecrypto formal verification blueprint — post-quantum ML-KEM/ML-DSA in iMessage
• 028. OpenClaw 安全和功能增强实践
• 029. xz-utils Backdoor 2 Years On — Maintainer Trust Hijack Pattern Beyond CVE Scanners
• 030. Disgruntled researcher releases two more Microsoft zero-days
• 031. Mythos 对企业安全架构影响的思考
• 032. Where OpenClaw Security Is Heading — OpenClaw Blog
• 033. Disgruntled researcher releases two more Microsoft zero-days
• 034. Canvas Breach Disrupts Schools & Colleges Nationwide
• 035. Adversaries Leverage AI for Vulnerability Exploitation, Augmented Operations, and Initial Access
• 036. 别让你的 Amazon Bedrock 模型为他人打工——API 调用安全防护指南
• 037. 首篇VLA安全综述出炉：机器人听懂指令，安全评测也要升级
• 038. Postmortem: TanStack npm supply-chain compromise | TanStack Blog
• 039. Canvas Hackers ShinyHunters Say Their Official Domain Was Suspended
• 040. 100万+AI服务暴露在公网——HackerNews扫描报告
• 041. Optimize blueprint extraction accuracy in Amazon Bedrock Data Automation
• 042. Disgruntled researcher releases two more Microsoft zero-days
• 043. How Amazon Bedrock catches AI-generated phishing
• 044. GitLab CI/CD Kill Chain Audit — Black Hills InfoSec 2026 大规模审计研究
• 045. Cyberscammers are bypassing banks' security with illicit tools sold on Telegram
• 046. INTERPOL Operation Ramz MENA Cybercrime Networks
• 047. AI in Cybersecurity Training Resources | SANS Institute
• 048. Canvas LMS 攻击者 ShinyHunters 官方域名被暂停：转向暗网的运营安全转向
• 049. U of T AI Worm：CleverHans Lab 展示可自适应的 AI 蠕虫威胁
• 050. Securing AI Agents and Machine Identities
• 051. A 0-click exploit chain for the Pixel 10: When a Door Closes, a Window Opens
• 052. ICO fines South Staffordshire £963K over 2022 breach
• 053. NGINX Rift: Achieving NGINX Remote Code Execution via an 18-Year-Old Vulnerability | depthfirst
• 054. ICO 对 South Staffordshire 处以 96.3 万英镑罚款：2022 年 Cl0p 勒索软件攻击暴露的安全失败
• 055. Fake Job Interview Apps Drop JobStealer Malware on Windows and macOS
• 056. bagel — Fleet 级 Secret Scanning 守护开发工作站
• 057. ICO fines Cl0p victim South Staffs Water over data breach
• 058. A 0-click exploit chain for the Pixel 10: When a Door Closes, a Window Opens
• 059. Pwn2Own Berlin 2026, Day Three: DEVCORE Crowned Master of Pwn, $1.298 Million Total
• 060. Funnel Builder Flaw Under Active Exploitation Enables WooCommerce Checkout Skimming
• 061. Autonomous Vulnerability Hunting with MCP
• 062. Sandworm Hackers Shift From IT Breaches to Critical OT Targets
• 063. GlassWASM: WebAssembly Malware Found in Trojanized Open VSX Extensions
• 064. Static Devirtualization of Themida
• 065. NIST SP 800-213r1 — IoT Product Cybersecurity Guidelines
• 066. CVE-2026-20182: Unauthenticated Cisco SD-WAN Control Plane Compromise via vHub Authentication Bypass
• 067. Stealing Passwords via HTML Injection Under a Strict CSP
• 068. TeamPCP Claims Sale of Mistral AI Repositories Amid Mini Shai-Hulud Attack
• 069. Meta U-turns on encryption push for Instagram as DMs go plaintext
• 070. OpenAI launches Daybreak to combat cyber threats
• 071. AI Voice Cloning: The Technology Behind It, Who's Building It, and Where It's Headed
• 072. Grafana GitHub Token Breach Led to Codebase Download and Extortion Attempt
• 073. Temporarily disabling new user registrations
• 074. Google and Amnesty International teamed up to make Android spyware detectable
• 075. RFC 9958: Post-Quantum Cryptography for Engineers
• 076. Thinkst Package Proxy: Supply Chain Safety Checks
• 077. Disgruntled researcher releases two more Microsoft zero-days
• 078. Funnel Builder 漏洞正被利用于 WooCommerce 支付 skimming
• 079. The down fall of bug bounties
• 080. Mozilla warns UK: Breaking VPNs will not magically fix Britain's age-check mess
• 081. JetBrains Marketplace Ecosystem Security Update: Malicious AI Plugins
• 082. LLMReaper: 浏览器扩展的对话窃取攻击
• 083. AI Detection and Response Aidr a Zero Impact Operating Model
• 084. Exploiting vulnerabilities in Johnson & Johnson web apps
• 085. Romanian Man Faces Up to 30 Years in US Prison Over Vishing Scams
• 086. Milvus 3.0 聚合下推：3 层隔离、2 条路径、4 倍安全因子
• 087. GitHub Breached — Employee Device Hack Led to Exfiltration of 3,800+ Internal Repos
• 088. Hackers accessed BWH Hotels reservation system for months
• 089. Getting a CVE Without Shipping Slop
• 090. The IT and security field guide to AI adoption | Tines
• 091. Unlocking the Cloudflare app ecosystem with OAuth for all
• 092. Mythos for Offensive Security: XBOW's Evaluation
• 093. Fedora Hummingbird brings the container security model to a Linux host OS
• 094. ShinyHunters hack 7-Eleven: franchisee data and Salesforce records exposed
• 095. GitHub Secret Scanning: AI/ML 驱动的大规模误报降低
• 096. 飞连 AI 办公安全体系 1+3+N 架构
• 097. Meet Bluekit: The AI-Powered All-in-One Phishing Kit
• 098. Building is just the beginning: Introducing Discoverability | Lovable
• 099. Scammers Send Physical Phishing Letters to Steal Ledger Wallet Seed Phrases
• 100. Cyberscammers are bypassing banks' security with illicit tools sold on Telegram
• 101. How Unified EDR and ITDR Stop Attacks Before They Spread
• 102. Guide to Security Operations at Machine Speed
• 103. Discord 全平台端到端加密
• 104. Nikesh Arora 20VC 访谈：Token 定价、FDE、SaaS→AI 转型与记忆护城河
• 105. Funnel Builder Flaw Under Active Exploitation Enables WooCommerce Checkout Skimming
• 106. Forward launches Predict to verify network changes before they reach production - SiliconANGLE
• 107. Static Devirtualization of Themida
• 108. CyberSecQwen-4B
• 109. Mystery Microsoft bug leaker keeps the zero-days coming
• 110. GitHub Breached — Employee Device Hack Led to Exfiltration
• 111. AI phishing attacks are on the rise — Are you prepared? | Bitwarden
• 112. cPanel, WHM Release Fixes for Three New Vulnerabilities — Patch Now
• 113. Semgrep Intercom Php Security
• 114. peerd: 浏览器原生的 AI Agent Harness
• 115. A DOD contractor’s API flaw exposed military course data and service member records
• 116. OpenSandbox：阿里开源的云端 Agent 安全沙箱（凭据 Vault + egress sidecar）
• 117. Anthropic's bug-hunting Mythos was greatest marketing stunt ever, says cURL creator
• 118. Incendium Fuzzing Ms Rpc
• 119. How Semgrep Cut Taint Analysis Time by 75%
• 120. On Post-Quantum Security Adoption
• 121. Jane Street — 形式化方法与编程的未来
• 122. Semgrep Intercom Php Supply Chain
• 123. 中国用户安全高性能访问海外 Bedrock
• 124. Drupal to Release Urgent Core Security Updates on May 20, Sites Told to Prepare
• 125. Cyberscammers are bypassing banks’ security with illicit tools sold on Telegram
• 126. DeepSecBench：评估 AI 模型在网络安全漏洞发现中的性能
• 127. Agent 审计：从海量噪音中捞出真风险
• 128. A 0-click exploit chain for the Pixel 10: When a Door Closes, a Window Opens
• 129. U of T researchers demonstrate AI worm: self-spreading malware using open-weight models
• 130. Japan’s PM orders cybersecurity review to defend against Anthropic Mythos