{"schema_version":"1.0","language":"zh-CN","generated_from":"NEWVAR 从零开始学AI structured Markdown","write_access":false,"full_text_access":false,"content_boundary":"This endpoint exposes metadata, objectives, prerequisites, labs and source summaries. It does not expose chapter body text or answer text.","chapter":{"slug":"agent-evaluation-security","title":"Agent评估、安全与人在回路","english_title":"Agent evaluation and security","chapter_number":84,"volume":4,"volume_title":"第四卷：前沿、系统与研究","month":21,"unit_title":"多模态、世界模型与Agent","review_status":"draft","updated_at":"2026-08-06","estimated_hours":10,"focus":"测试任务成功、引用、权限、提示注入、工具故障和可恢复性。","objectives":["测试任务成功、引用、权限、提示注入、工具故障和可恢复性。","解释核心数学关系并完成最小实现","运行单变量实验并记录失败样例","完成工程线或研究线至少一项迁移任务"],"prerequisites":["agent-loop-tools-memory"],"exercise_count":3,"tracks":["engineering","research"],"human_url":"/learn/agent-evaluation-security"},"labs":[{"id":"lab-agent-evaluation-security","chapter_slug":"agent-evaluation-security","title":"Agent评估、安全与人在回路：单变量实验","kind":"autonomy_level","data_status":"教学模拟","variable":{"label":"自主级别","unit":"级","min":0,"max":5,"step":1,"default":1},"objective":"只改变“自主级别”，观察结果、代价和风险如何一起变化，理解测试任务成功、引用、权限、提示注入、工具故障和可恢复性。","static_fallback":"静态替代：把自主级别分别设为0级、1级和5级，手工比较三组教学模拟输出。","keyboard":"聚焦滑块后使用方向键微调，Page Up/Page Down大步调整，Home/End到达边界。","measurement_note":"页面数值由公开公式生成，只用于教学，不代表真实模型性能或现实世界因果效果。","source_ids":["S056","S057","S058","S059","S060","S084","S085"]}],"project":{"id":"project-21","month":21,"title":"做一个只读、可审计的研究助理Agent","brief":"限制工具白名单，记录每次检索和引用，遇到发布、写入或高风险判断必须停下等待人工确认。","deliverables":["问题与边界说明","可运行最小实现","实验记录与失败分析","来源与许可清单","风险、隐私与人工闸门","复现README"]},"sources":[{"id":"S056","author":"Alec Radford et al.","title":"Learning Transferable Visual Models From Natural Language Supervision","source_level":"同行评审论文预印本","published_at":"2021","accessed_at":"2026-08-06","doi_or_url":"https://arxiv.org/abs/2103.00020"},{"id":"S057","author":"David Ha, Jürgen Schmidhuber","title":"World Models","source_level":"同行评审论文预印本","published_at":"2018","accessed_at":"2026-08-06","doi_or_url":"https://arxiv.org/abs/1803.10122"},{"id":"S058","author":"Shunyu Yao et al.","title":"ReAct: Synergizing Reasoning and Acting in Language Models","source_level":"同行评审论文预印本","published_at":"2022","accessed_at":"2026-08-06","doi_or_url":"https://arxiv.org/abs/2210.03629"},{"id":"S059","author":"Timo Schick et al.","title":"Toolformer","source_level":"同行评审论文预印本","published_at":"2023","accessed_at":"2026-08-06","doi_or_url":"https://arxiv.org/abs/2302.04761"},{"id":"S060","author":"Xiao Liu et al.","title":"AgentBench","source_level":"研究论文预印本","published_at":"2023","accessed_at":"2026-08-06","doi_or_url":"https://arxiv.org/abs/2308.03688"},{"id":"S084","author":"OWASP Foundation","title":"OWASP Top 10 for LLM Applications","source_level":"行业安全项目一手资料","published_at":"持续更新","accessed_at":"2026-08-06","doi_or_url":"https://genai.owasp.org/llm-top-10/"},{"id":"S085","author":"MITRE","title":"MITRE ATLAS","source_level":"威胁知识库一手资料","published_at":"持续更新","accessed_at":"2026-08-06","doi_or_url":"https://atlas.mitre.org/"}]}