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 113 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. A Framework for AI Threat Readiness
• 003. From SSH to REST: A Security-Driven Modernization of Slack's EMR Data Pipelines
• 004. Mozilla warns UK: Breaking VPNs will not magically fix Britain's age-check mess
• 005. Offensive Security Blog
• 006. Sandworm Hackers Shift From IT Breaches to Critical OT Targets
• 007. 5 Things to Know about the CLARITY Act
• 008. fedora hummingbird brings the container security model to a linux host os
• 009. Mythos finds a curl vulnerability
• 010. 飞来汇借助 AWS Security Agent 构建跨境支付应用的智能安全防线
• 011. Canvas Hackers ShinyHunters Say Their Official Domain Was Suspended
• 012. Hermes Agent v0.14.0 核心架构与快速上手
• 013. bleeding-llama-critical-unauthenticated-memory-leak-in-ollama
• 014. SHub Reaper: macOS Stealer Spoofs Apple, Google, and Microsoft in a Single Attack Chain
• 015. Resecurity | CVE-2026-20182: Unauthenticated Cisco SD-WAN Control-Plane Compromise via vHub Authentication Bypass
• 016. LLMReaper - DOM Based AI Conversation Exfiltration via Browser Extensions
• 017. Static Devirtualization of Themida
• 018. Static Devirtualization 2024
• 019. What My Privacy and Security Stack Actually Looks Like
• 020. How an image could compromise your
• 021. Alliance for Critical Infrastructure (ACI): US Critical Infrastructure Cybersecurity Coalition
• 022. Static Devirtualization of Themida
• 023. Inference Theft as AI Endpoint Attack Surface — Vercel Token Theft Defense 2026
• 024. Apple corecrypto formal verification blueprint — post-quantum ML-KEM/ML-DSA in iMessage
• 025. OpenClaw 安全和功能增强实践
• 026. Optimize blueprint extraction accuracy in Amazon Bedrock Data Automation"
• 027. xz-utils Backdoor 2 Years On — Maintainer Trust Hijack Pattern Beyond CVE Scanners
• 028. Disgruntled researcher releases two more Microsoft zero-days
• 029. Disgruntled researcher releases two more Microsoft zero-days
• 030. Where OpenClaw Security Is Heading — OpenClaw Blog
• 031. Canvas Breach Disrupts Schools & Colleges Nationwide
• 032. 别让你的 Amazon Bedrock 模型为他人打工——API 调用安全防护指南
• 033. Postmortem: TanStack npm supply-chain compromise | TanStack Blog
• 034. 100万+AI服务暴露在公网——HackerNews扫描报告
• 035. Canvas Hackers ShinyHunters Say Their Official Domain Was Suspended
• 036. Adversaries Leverage AI for Vulnerability Exploitation, Augmented Operations, and Initial Access
• 037. Disgruntled researcher releases two more Microsoft zero-days
• 038. GitLab CI/CD Kill Chain Audit — Black Hills InfoSec 2026 大规模审计研究
• 039. INTERPOL Operation Ramz MENA Cybercrime Networks
• 040. Cyberscammers are bypassing banks' security with illicit tools sold on Telegram
• 041. Securing AI Agents and Machine Identities
• 042. A 0-click exploit chain for the Pixel 10: When a Door Closes, a Window Opens
• 043. U of T AI Worm：CleverHans Lab 展示可自适应的 AI 蠕虫威胁
• 044. Canvas LMS 攻击者 ShinyHunters 官方域名被暂停：转向暗网的运营安全转向
• 045. NGINX Rift: Achieving NGINX Remote Code Execution via an 18-Year-Old Vulnerability | depthfirst
• 046. Fake Job Interview Apps Drop JobStealer Malware on Windows and macOS
• 047. 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
• 048. AI in Cybersecurity Training Resources | SANS Institute
• 049. bagel — Fleet 级 Secret Scanning 守护开发工作站
• 050. Funnel Builder Flaw Under Active Exploitation Enables WooCommerce Checkout Skimming
• 051. Pwn2Own Berlin 2026, Day Three: DEVCORE Crowned Master of Pwn, $1.298 Million Total
• 052. Autonomous Vulnerability Hunting with MCP
• 053. A 0-click exploit chain for the Pixel 10: When a Door Closes, a Window Opens
• 054. Static Devirtualization of Themida
• 055. TeamPCP Claims Sale of Mistral AI Repositories Amid Mini Shai-Hulud Attack
• 056. Grafana GitHub Token Breach Led to Codebase Download and Extortion Attempt
• 057. Meta U-turns on encryption push for Instagram as DMs go plaintext
• 058. AI Voice Cloning: The Technology Behind It, Who's Building It, and Where It's Headed
• 059. Google and Amnesty International teamed up to make Android spyware detectable
• 060. RFC 9958: Post-Quantum Cryptography for Engineers
• 061. Disgruntled researcher releases two more Microsoft zero-days
• 062. OpenAI launches Daybreak to combat cyber threats
• 063. Funnel Builder 漏洞正被利用于 WooCommerce 支付 skimming
• 064. The down fall of bug bounties
• 065. Mozilla warns UK: Breaking VPNs will not magically fix Britain's age-check mess
• 066. JetBrains Marketplace Ecosystem Security Update: Malicious AI Plugins
• 067. GlassWASM: WebAssembly Malware Found in Trojanized Open VSX Extensions
• 068. ai detection and response aidr a zero impact operating model
• 069. GitHub Breached — Employee Device Hack Led to Exfiltration of 3,800+ Internal Repos
• 070. Exploiting vulnerabilities in Johnson & Johnson web apps
• 071. The IT and security field guide to AI adoption | Tines
• 072. Mythos for Offensive Security: XBOW's Evaluation
• 073. Getting a CVE Without Shipping Slop
• 074. Fedora Hummingbird brings the container security model to a Linux host OS
• 075. Meet Bluekit: The AI-Powered All-in-One Phishing Kit
• 076. Building is just the beginning: Introducing Discoverability | Lovable
• 077. Scammers Send Physical Phishing Letters to Steal Ledger Wallet Seed Phrases
• 078. LLMReaper: 浏览器扩展的对话窃取攻击
• 079. How Unified EDR and ITDR Stop Attacks Before They Spread
• 080. Cyberscammers are bypassing banks' security with illicit tools sold on Telegram
• 081. Guide to Security Operations at Machine Speed
• 082. Discord 全平台端到端加密
• 083. A 0-click exploit chain for the Pixel 10: When a Door Closes, a Window Opens
• 084. Funnel Builder Flaw Under Active Exploitation Enables WooCommerce Checkout Skimming
• 085. Forward launches Predict to verify network changes before they reach production - SiliconANGLE
• 086. Static Devirtualization of Themida
• 087. GitHub Breached — Employee Device Hack Led to Exfiltration
• 088. CyberSecQwen-4B
• 089. Mystery Microsoft bug leaker keeps the zero-days coming
• 090. AI phishing attacks are on the rise — Are you prepared? | Bitwarden
• 091. cPanel, WHM Release Fixes for Three New Vulnerabilities — Patch Now
• 092. semgrep intercom php security
• 093. A DOD contractor’s API flaw exposed military course data and service member records
• 094. Anthropic's bug-hunting Mythos was greatest marketing stunt ever, says cURL creator
• 095. incendium fuzzing ms rpc
• 096. How Semgrep Cut Taint Analysis Time by 75%
• 097. Sandworm Hackers Shift From IT Breaches to Critical OT Targets
• 098. On Post-Quantum Security Adoption
• 099. Jane Street — 形式化方法与编程的未来
• 100. semgrep intercom php supply chain
• 101. ICO fines Cl0p victim South Staffs Water over data breach
• 102. peerd: 浏览器原生的 AI Agent Harness
• 103. Drupal to Release Urgent Core Security Updates on May 20, Sites Told to Prepare
• 104. Cyberscammers are bypassing banks’ security with illicit tools sold on Telegram
• 105. Hackers accessed BWH Hotels reservation system for months
• 106. ICO fines South Staffordshire £963K over 2022 breach
• 107. 中国用户安全高性能访问海外 Bedrock
• 108. ShinyHunters hack 7-Eleven: franchisee data and Salesforce records exposed
• 109. Temporarily disabling new user registrations
• 110. Romanian Man Faces Up to 30 Years in US Prison Over Vishing Scams
• 111. GitHub Secret Scanning: AI/ML 驱动的大规模误报降低
• 112. U of T researchers demonstrate AI worm: self-spreading malware using open-weight models
• 113. Japan’s PM orders cybersecurity review to defend against Anthropic Mythos