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 58 篇实体，按深度递增排列。
本章导航¶
| Level | 含义 | 篇数 |
|---|---|---|
| ⭐ 入门 | 零基础可读 | 3 |
| ⭐⭐ 工程师 | 需编程基础 | 5 |
| ⭐⭐⭐ 专家 | 需ML基础 | 26 |
| ⭐⭐⭐⭐ 科学家 | 需研究背景 | 19 |
| ⭐⭐⭐⭐⭐ 大师 | 前沿/哲学 | 5 |
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¶
⭐ 入门（3 篇）¶
- 001. Malware crew TeamPCP open-sources its Shai-Hulud worm on GitHub
- 002. peerd: 浏览器原生的 AI Agent Harness
- 003. Jane Street — 形式化方法与编程的未来
⭐⭐ 工程师（5 篇）¶
- 004. OpenAI launches Daybreak to combat cyber threats
- 005. Grafana GitHub Token Breach Led to Codebase Download and Extortion Attempt
- 006. JetBrains Marketplace Ecosystem Security Update: Malicious AI Plugins
- 007. Unlocking the Cloudflare app ecosystem with OAuth for all
- 008. LLMShare: using shared chatbot pages to distribute malware
⭐⭐⭐ 专家（26 篇）¶
- 009. 飞来汇借助 AWS Security Agent 构建跨境支付应用的智能安全防线
- 010. CISA urges critical infrastructure firms to 'fortify' before it's too late | Cybersecurity Dive
- 011. Disgruntled researcher releases two more Microsoft zero-days
- 012. Where OpenClaw Security Is Heading — OpenClaw Blog
- 013. Adversaries Leverage AI for Vulnerability Exploitation, Augmented Operations, and Initial Access
- 014. 别让你的 Amazon Bedrock 模型为他人打工——API 调用安全防护指南
- 015. Canvas Hackers ShinyHunters Say Their Official Domain Was Suspended
- 016. A Framework for AI Threat Readiness
- 017. Cyberscammers are bypassing banks' security with illicit tools sold on Telegram
- 018. How Semgrep Cut Taint Analysis Time by 75%
- 019. U of T AI Worm：CleverHans Lab 展示可自适应的 AI 蠕虫威胁
- 020. 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
- 021. NGINX Rift: Achieving NGINX Remote Code Execution via an 18-Year-Old Vulnerability | depthfirst
- 022. Offensive Security Blog
- 023. Pwn2Own Berlin 2026, Day Three: DEVCORE Crowned Master of Pwn, $1.298 Million Total
- 024. TeamPCP Claims Sale of Mistral AI Repositories Amid Mini Shai-Hulud Attack
- 025. AI Voice Cloning: The Technology Behind It, Who's Building It, and Where It's Headed
- 026. The down fall of bug bounties
- 027. Exploiting vulnerabilities in Johnson & Johnson web apps
- 028. GitHub Breached — Employee Device Hack Led to Exfiltration of 3,800+ Internal Repos
- 029. Fedora Hummingbird brings the container security model to a Linux host OS
- 030. Guide to Security Operations at Machine Speed
- 031. Forward launches Predict to verify network changes before they reach production - SiliconANGLE
- 032. GitHub Breached — Employee Device Hack Led to Exfiltration
- 033. AI phishing attacks are on the rise — Are you prepared? | Bitwarden
- 034. CyberSecQwen-4B
⭐⭐⭐⭐ 科学家（19 篇）¶
- 035. Mythos finds a curl vulnerability
- 036. LLMReaper - DOM Based AI Conversation Exfiltration via Browser Extensions
- 037. Resecurity | CVE-2026-20182: Unauthenticated Cisco SD-WAN Control-Plane Compromise via vHub Authentication Bypass
- 038. Static Devirtualization of Themida
- 039. Apple corecrypto formal verification blueprint — post-quantum ML-KEM/ML-DSA in iMessage
- 040. OpenClaw 安全和功能增强实践
- 041. Nikesh Arora 20VC 访谈：Token 定价、FDE、SaaS→AI 转型与记忆护城河
- 042. 中国用户安全高性能访问海外 Bedrock
- 043. xz-utils Backdoor 2 Years On — Maintainer Trust Hijack Pattern Beyond CVE Scanners
- 044. Token 撤销触发设备擦除的安全漏洞
- 045. 100万+AI服务暴露在公网——HackerNews扫描报告
- 046. Optimize blueprint extraction accuracy in Amazon Bedrock Data Automation
- 047. GitLab CI/CD Kill Chain Audit — Black Hills InfoSec 2026 大规模审计研究
- 048. A 0-click exploit chain for the Pixel 10: When a Door Closes, a Window Opens
- 049. From SSH to REST: A Security-Driven Modernization of Slack's EMR Data Pipelines
- 050. bagel — Fleet 级 Secret Scanning 守护开发工作站
- 051. GlassWASM: WebAssembly Malware Found in Trojanized Open VSX Extensions
- 052. Getting a CVE Without Shipping Slop
- 053. GitHub Secret Scanning: AI/ML 驱动的大规模误报降低
⭐⭐⭐⭐⭐ 大师（5 篇）¶
- 054. Hermes Agent v0.14.0 核心架构与快速上手
- 055. Bleeding Llama：Ollama 未授权内存泄漏漏洞
- 056. SHub Reaper: macOS Stealer Spoofs Apple, Google, and Microsoft in a Single Attack Chain
- 057. How an image could compromise your
- 058. Securing AI Agents and Machine Identities
本章收束¶
权限越大，攻击面越大。这一章的案例库——凭据泄露、提示注入、供应链投毒——指向同一条设计原则：安全不是上线前的检查项，而是架构第一天的承重墙。给 Agent 权限之前，先给它身份、边界和审计；这三样齐了，出事才知道坏在哪、止得住。