全书目录

附录 G：分章资料清单

G.54 第 27 章正文采用的核心来源

2 分钟 662 字 第 110 / 161 个阅读单元
  • 官方注入与工具误用主轴：OpenAI Safety in building agents / Prompt injection best practices / Designing AI agents to resist prompt injection / Operator System Card
  • 分层防御与最小权限主轴：Anthropic Mitigate jailbreaks and prompt injections / Securely deploying AI agents / Computer use、OpenAI Instruction Hierarchy
  • 企业运行时防护主轴：Google Model Armor / AI security and safety、Microsoft Prompt Shields / Prompt Injection Attacks
  • 协议与策略控制主轴：MCP Security Best Practices、OWASP LLM01 Prompt Injection / LLM08 Excessive Agency
  • 状态污染与攻击链分析主轴：Make Agent Defeat Agent、PoisonedRAG
  • Benchmark 与本章落点主轴：Indirect Prompt Injection、AgentDojo、Agent Security Bench、AI Agents Under Threat，共同支撑“这四类风险本质上是一条可串联的自动化失效链”