5 天 Agent 安全
课程题库 · 15 / 15 道题
D1 威胁模型:致命三件套、信任边界,以及两张 OWASP 清单怎么当检查表用
什么是致命三件套?为什么说它描述的是组合风险而不是单点缺陷?What is the lethal trifecta, and why is it a compositional risk rather than a single-point defect?
国内高频海外高频基础#threat-modeling#lethal-trifecta🔒 解锁 D1 后查看答题要点。
直接注入和间接注入有什么区别?哪一种更难防,为什么?What is the difference between direct and indirect prompt injection, and which is harder to defend against?
国内高频海外高频进阶#prompt-injection#attack-surface🔒 解锁 D1 后查看答题要点。
给你一个已经上线的 Agent,你怎么在半小时内画出它的信任边界并找出最高危的那条链路?Given an agent already in production, how would you map its trust boundary and find its highest-risk path in half an hour?
国内高频海外高频深入#threat-modeling#trust-boundary🔒 解锁 D1 后查看答题要点。
D2 注入靶场与输入侧防御:四档防御各自挡住了什么,又各自在哪里失手
一个 Agent 的注入面一共有几条?工具返回的内容算不算不可信输入?How many injection surfaces does an agent have, and do tool results count as untrusted input?
国内高频海外高频基础#prompt-injection#threat-model🔒 解锁 D2 后查看答题要点。
为什么说提示词里的分隔符和检测器都不能当作安全边界?它们各自能做到什么程度?Why are prompt delimiters and detectors not security boundaries, and how far does each actually get you?
国内高频海外高频进阶#prompt-injection#input-defense#adaptive-attack🔒 解锁 D2 后查看答题要点。
评估一个注入防御方案时,只看攻击成功率会漏掉什么?你会怎么设计这个评估?What does an evaluation miss if it only reports attack success rate, and how would you design it instead?
国内高频海外高频深入#evaluation#prompt-injection#metrics🔒 解锁 D2 后查看答题要点。
D3 架构级防御:六个设计模式,以及先定计划后执行怎么把攻击成功率压到零
先定计划后执行为什么能挡住间接注入?它挡不住什么?Why does plan-then-execute stop indirect prompt injection, and what does it fail to stop?
国内高频海外高频进阶#prompt-injection#architecture#plan-then-execute🔒 解锁 D3 后查看答题要点。
双模型隔离和加一个检测器,本质区别在哪?What is the fundamental difference between dual-LLM isolation and adding a detector?
国内高频海外高频深入#dual-llm#detector#threat-modeling🔒 解锁 D3 后查看答题要点。
什么情况下你会拒绝使用这些模式,宁可接受风险?When would you decline to apply these patterns and accept the risk instead?
国内高频海外高频基础#risk-tradeoff#lethal-trifecta#architecture🔒 解锁 D3 后查看答题要点。
D4 运行时管控:能力面、确认门、出口白名单、沙箱分层,以及密钥与租户隔离
人在回路的确认门,为什么必须放在服务端的工具执行器里?放在前端会出什么事?Why must a human-in-the-loop confirmation gate live inside the server-side tool executor? What breaks if you put it in the frontend?
国内高频海外高频进阶#human-in-the-loop#tool-execution#trust-boundary🔒 解锁 D4 后查看答题要点。
网络出口白名单能挡住哪些数据外泄路径?哪些是它挡不住的?Which data exfiltration paths does a network egress allowlist actually block, and which ones does it miss?
国内高频海外高频深入#egress-control#ssrf#data-exfiltration🔒 解锁 D4 后查看答题要点。
一个多租户的 Agent 服务里,凭据和日志分别该怎么隔离?In a multi-tenant agent service, how should credentials and logs be isolated?
国内高频海外高频进阶#multi-tenancy#secrets-management#audit-logging🔒 解锁 D4 后查看答题要点。
D5 红队与演练:攻击用例集怎么长期维护、怎么进 CI,以及出事之后怎么办
红队测试进 CI 之后,阈值该怎么定才既有效又不会天天误报?Once red-teaming runs in CI, how do you set thresholds that are both meaningful and not a daily false alarm?
国内高频海外高频深入#red-teaming#ci-gating🔒 解锁 D5 后查看答题要点。
怎么判断一个 Agent 已经失控?断路开关该由谁来拉?How do you tell an agent has gone rogue, and who gets to pull the kill switch?
国内高频海外高频进阶#rogue-agent#kill-switch🔒 解锁 D5 后查看答题要点。
一次 Agent 数据外泄事故发生后,你的前三步分别做什么?After an agent data exfiltration incident, what are your first three steps?
国内高频海外高频进阶#incident-response#forensics🔒 解锁 D5 后查看答题要点。