{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# 第10章 学生Notebook：受控学习便签助手\n", "\n", "> 模型没有直接写文件的魔法——工具调用是\"提议\"，宿主程序才\"执行\"\n", "\n", "本Notebook完成以下任务：\n", "1. 加载工具 Schema，理解参数约束\n", "2. 运行正常工具调用流程（创建笔记 → 验证存在）\n", "3. 对比\"模型声称已创建\"与\"日志证明已执行\"\n", "4. 测试5种攻击场景，观察防御机制\n", "5. 查看审计日志\n", "6. B档修改：修改一条 Schema 限制或新增一个低风险工具\n", "\n", "**准备**：确认 `data/tool_schemas.json` 存在。" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 1. 加载工具 Schema\n", "\n", "读取三个工具的定义，理解每个工具的参数约束和权限。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import json\n", "import re\n", "import os\n", "import tempfile\n", "import shutil\n", "from pathlib import Path\n", "from datetime import datetime\n", "\n", "# 加载工具 Schema\n", "schema_path = Path('../data/tool_schemas.json')\n", "with open(schema_path, 'r', encoding='utf-8') as f:\n", " schema_data = json.load(f)\n", "\n", "tools = {t['name']: t for t in schema_data['tools']}\n", "global_rules = schema_data['global_rules']\n", "\n", "print(f'工具数量: {len(tools)}')\n", "print(f'全局规则: {json.dumps(global_rules, ensure_ascii=False, indent=2)}')\n", "print()\n", "for name, tool in tools.items():\n", " print(f'工具: {name}')\n", " print(f' 说明: {tool[\"description\"]}')\n", " print(f' 权限: {tool[\"permissions\"]}')\n", " print(f' 风险: {tool[\"risk_level\"]}')\n", " params = tool['parameters'].get('properties', {})\n", " if params:\n", " for pname, pdef in params.items():\n", " print(f' 参数 {pname}: {pdef.get(\"type\", \"?\")}', end='')\n", " if 'pattern' in pdef:\n", " print(f', 正则: {pdef[\"pattern\"]}', end='')\n", " if 'maxLength' in pdef:\n", " print(f', 最大长度: {pdef[\"maxLength\"]}', end='')\n", " print()\n", " print()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 2. 搭建沙盒与审计日志\n", "\n", "创建临时沙盒目录和审计日志。所有文件操作都限制在这个目录中。\n", "\n", "**运行**：观察沙盒路径。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# 创建沙盒目录（使用临时目录，安全隔离）\n", "SANDBOX_ROOT = Path(tempfile.mkdtemp(prefix='ch10_sandbox_'))\n", "AUDIT_LOG = [] # 内存中的审计日志\n", "\n", "print(f'沙盒目录: {SANDBOX_ROOT}')\n", "print(f'沙盒目录存在: {SANDBOX_ROOT.exists()}')\n", "print()\n", "print('安全说明:')\n", "print(' - 所有文件操作限制在沙盒目录内')\n", "print(' - 不访问沙盒外的任何文件')\n", "print(' - 每次调用都记录审计日志')\n", "print(' - 课程结束后可安全删除整个沙盒')" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 3. 实现校验层\n", "\n", "这是本章的核心代码：**模型只提议，程序来校验**。\n", "\n", "校验分为四层：Schema 校验 → 路径安全检查 → 内容安全检查 → 权限检查。\n", "\n", "**运行**：阅读完整代码，理解每一层检查什么。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "def validate_schema(tool_name, arguments):\n", " \"\"\"第一层：Schema 校验 —— 参数类型、格式、长度是否符合工具定义\"\"\"\n", " if tool_name not in tools:\n", " return False, f'unknown_tool: {tool_name}'\n", " \n", " tool = tools[tool_name]\n", " schema = tool['parameters']\n", " \n", " # 检查必填参数\n", " for req in schema.get('required', []):\n", " if req not in arguments:\n", " return False, f'missing_required: {req}'\n", " \n", " # 检查额外属性\n", " if schema.get('additionalProperties') is False:\n", " allowed = set(schema.get('properties', {}).keys())\n", " extra = set(arguments.keys()) - allowed\n", " if extra:\n", " return False, f'additional_properties: {extra}'\n", " \n", " # 检查每个参数的类型和格式\n", " properties = schema.get('properties', {})\n", " for pname, pvalue in arguments.items():\n", " if pname not in properties:\n", " continue\n", " pdef = properties[pname]\n", " \n", " # 类型检查\n", " if pdef.get('type') == 'string' and not isinstance(pvalue, str):\n", " return False, f'type_error: {pname} should be string'\n", " \n", " # 正则检查\n", " if 'pattern' in pdef and isinstance(pvalue, str):\n", " if not re.match(pdef['pattern'], pvalue):\n", " return False, f'pattern_error: {pname} does not match {pdef[\"pattern\"]}'\n", " \n", " # 长度检查\n", " if 'maxLength' in pdef and isinstance(pvalue, str):\n", " if len(pvalue) > pdef['maxLength']:\n", " return False, f'maxLength_error: {pname} is {len(pvalue)} chars, max {pdef[\"maxLength\"]}'\n", " \n", " return True, 'schema_ok'\n", "\n", "\n", "def validate_path_safety(arguments):\n", " \"\"\"第二层：路径安全检查 —— 防止路径穿越\"\"\"\n", " for key, value in arguments.items():\n", " if not isinstance(value, str):\n", " continue\n", " for blocked in ['..', '/', '\\\\']:\n", " if blocked in value:\n", " return False, f'path_traversal: \"{blocked}\" found in {key}'\n", " return True, 'path_ok'\n", "\n", "\n", "def validate_content_safety(arguments):\n", " \"\"\"第三层：内容安全检查 —— 防止文档注入\"\"\"\n", " for key, value in arguments.items():\n", " if not isinstance(value, str):\n", " continue\n", " value_lower = value.lower()\n", " for blocked in global_rules['blocked_patterns']:\n", " if blocked.lower() in value_lower:\n", " return False, f'blocked_pattern: \"{blocked}\" found in {key}'\n", " return True, 'content_ok'\n", "\n", "\n", "def validate_all(tool_name, arguments):\n", " \"\"\"依次运行四层校验，任何一层失败则拒绝\"\"\"\n", " checks = [\n", " ('schema', validate_schema(tool_name, arguments)),\n", " ('path_safety', validate_path_safety(arguments)),\n", " ('content_safety', validate_content_safety(arguments)),\n", " ]\n", " \n", " for layer_name, (passed, reason) in checks:\n", " if not passed:\n", " return False, layer_name, reason\n", " \n", " return True, 'all_passed', 'all_ok'\n", "\n", "\n", "print('校验层已定义:')\n", "print(' 1. validate_schema — 参数类型、格式、长度、额外属性')\n", "print(' 2. validate_path_safety — 路径穿越（.. / \\\\）')\n", "print(' 3. validate_content_safety — 文档注入（'\n", " }\n", " }\n", " },\n", " {\n", " 'description': '攻击：危险计算',\n", " 'tool_call': {\n", " 'tool': 'calc',\n", " 'arguments': {\n", " 'expression': \"__import__('os').system('echo hacked')\"\n", " }\n", " }\n", " }\n", "]\n", "\n", "print(f'预录模型输出数量: {len(mock_model_outputs)}')\n", "print()\n", "for i, item in enumerate(mock_model_outputs):\n", " tc = item['tool_call']\n", " print(f' [{i+1}] {item[\"description\"]}')\n", " print(f' tool={tc[\"tool\"]}', end='')\n", " if tc['arguments']:\n", " args_preview = {k: (v[:30]+'...' if isinstance(v,str) and len(v)>30 else v)\n", " for k, v in tc['arguments'].items()}\n", " print(f' args={args_preview}')\n", " else:\n", " print(' args={}')" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 6. 正常调用流程\n", "\n", "运行前三个正常调用，观察完整的\"提议→校验→执行→日志\"四段证据。\n", "\n", "**运行**：观察每次调用的结果。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "print('=' * 60)\n", "print('正常调用流程')\n", "print('=' * 60)\n", "\n", "for item in mock_model_outputs[:3]:\n", " tc = item['tool_call']\n", " print(f'\\n--- {item[\"description\"]} ---')\n", " print(f'提议: tool={tc[\"tool\"]}, args={tc[\"arguments\"]}')\n", " \n", " response = process_tool_call(tc['tool'], tc['arguments'])\n", " log = response['log']\n", " \n", " print(f'校验: passed={log[\"stages\"][\"validation\"][\"passed\"]}, '\n", " f'layer={log[\"stages\"][\"validation\"][\"layer\"]}')\n", " print(f'执行: {log[\"stages\"][\"execution\"][\"status\"]}')\n", " if response['result']:\n", " print(f'结果: {response[\"result\"]}')\n", " print(f'最终: {log[\"final\"]}')" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 验证文件确实存在\n", "\n", "模型说\"我已经创建了文件\"，但我们要用 `list_files` 工具来**实际验证**。\n", "\n", "**运行**：对比\"模型声称\"与\"日志证明\"。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "print('=' * 60)\n", "print('\"模型声称\" vs \"日志证明\"')\n", "print('=' * 60)\n", "\n", "# 模型声称\n", "print('\\n模型声称: \"我已经创建了今日清单.md\"')\n", "print(' → 这只是一段文字，模型没有文件操作能力')\n", "\n", "# 日志证明\n", "print('\\n日志证明:')\n", "for entry in AUDIT_LOG:\n", " if entry['tool'] == 'create_note' and entry['final'] == 'EXECUTED':\n", " print(f' [{entry[\"timestamp\"]}] create_note '\n", " f'filename={entry[\"arguments\"][\"filename\"]} → EXECUTED')\n", "\n", "# 实际验证\n", "print('\\n实际验证（调用 list_files）:')\n", "response = process_tool_call('list_files', {})\n", "if response['result']:\n", " print(f' 沙盒中的文件: {response[\"result\"][\"files\"]}')\n", " print(f' 文件数量: {response[\"result\"][\"count\"]}')\n", "\n", "# 直接用文件系统验证\n", "print('\\n文件系统直接验证:')\n", "actual_files = [f.name for f in SANDBOX_ROOT.iterdir() if f.is_file()]\n", "print(f' 沙盒目录实际文件: {actual_files}')\n", "print(f' \"今日清单.md\" 存在: {(\"今日清单.md\" in actual_files)}')\n", "\n", "print('\\n结论: 只有审计日志 + 文件系统验证才能证明操作确实执行。')\n", "print(' 模型的文字输出不等于执行证据。')" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "**思考**：\n", "\n", "如果模型说\"我已经创建了文件\"，但审计日志中没有对应记录，可能是什么原因？\n", "\n", "在下方写下你的分析。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# 我的分析：\n", "# \n", "# 可能原因1：\n", "# 可能原因2：\n", "# 可能原因3：" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 7. 攻击测试\n", "\n", "运行5种攻击场景，观察每层防御如何拦截非法调用。\n", "\n", "**运行**：观察每种攻击被哪一层防御拦截。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "print('=' * 60)\n", "print('攻击测试')\n", "print('=' * 60)\n", "\n", "attack_results = []\n", "\n", "for item in mock_model_outputs[3:]: # 从第4个开始是攻击\n", " tc = item['tool_call']\n", " print(f'\\n--- {item[\"description\"]} ---')\n", " \n", " # 显示攻击输入（截断长内容）\n", " args_display = {}\n", " for k, v in tc['arguments'].items():\n", " if isinstance(v, str) and len(v) > 60:\n", " args_display[k] = v[:60] + f'... ({len(v)} chars)'\n", " else:\n", " args_display[k] = v\n", " print(f'攻击输入: tool={tc[\"tool\"]}, args={args_display}')\n", " \n", " response = process_tool_call(tc['tool'], tc['arguments'])\n", " log = response['log']\n", " \n", " validation = log['stages']['validation']\n", " execution = log['stages']['execution']\n", " \n", " print(f'校验结果: passed={validation[\"passed\"]}')\n", " if not validation['passed']:\n", " print(f'拦截层: {validation[\"layer\"]}')\n", " print(f'拦截原因: {validation[\"reason\"]}')\n", " print(f'执行状态: {execution[\"status\"]}')\n", " print(f'最终: {log[\"final\"]}')\n", " \n", " attack_results.append({\n", " 'attack': item['description'],\n", " 'blocked': not validation['passed'],\n", " 'layer': validation.get('layer', 'N/A'),\n", " 'reason': validation.get('reason', 'N/A')\n", " })\n", "\n", "print('\\n' + '=' * 60)\n", "print('攻击防御汇总')\n", "print('=' * 60)\n", "print(f'{\"攻击类型\":<20} {\"拦截\":<6} {\"防御层\":<18} {\"原因\"}')\n", "print('-' * 80)\n", "for r in attack_results:\n", " status = '✓ 拦截' if r['blocked'] else '✗ 放行'\n", " print(f'{r[\"attack\"]:<20} {status:<6} {r[\"layer\"]:<18} {r[\"reason\"]}')" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "**思考**：\n", "\n", "1. 路径穿越攻击被哪一层拦截？如果没有这一层会怎样？\n", "2. 文档注入攻击被哪一层拦截？为什么内容安全检查很重要？\n", "3. 未声明字段攻击被哪一层拦截？`additionalProperties: false` 的作用是什么？\n", "\n", "在下方写下你的分析。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# 我的分析：\n", "# \n", "# 1. 路径穿越：\n", "# \n", "# 2. 文档注入：\n", "# \n", "# 3. 未声明字段：" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 8. 审计日志查看器\n", "\n", "查看完整的审计日志。每条记录包含：时间戳、工具名、参数摘要、校验结果、执行结果。\n", "\n", "**运行**：查看所有调用记录。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "print('=' * 60)\n", "print('完整审计日志')\n", "print('=' * 60)\n", "\n", "for i, entry in enumerate(AUDIT_LOG):\n", " print(f'\\n--- 记录 #{i+1} ---')\n", " print(f'时间: {entry[\"timestamp\"]}')\n", " print(f'工具: {entry[\"tool\"]}')\n", " print(f'参数: {entry[\"arguments\"]}')\n", " print(f'校验: passed={entry[\"stages\"][\"validation\"][\"passed\"]}, '\n", " f'layer={entry[\"stages\"][\"validation\"][\"layer\"]}')\n", " print(f'执行: {entry[\"stages\"][\"execution\"][\"status\"]}')\n", " print(f'最终: {entry[\"final\"]}')\n", "\n", "# 统计\n", "executed = sum(1 for e in AUDIT_LOG if e['final'] == 'EXECUTED')\n", "rejected = sum(1 for e in AUDIT_LOG if e['final'] == 'REJECTED')\n", "print(f'\\n--- 统计 ---')\n", "print(f'总调用次数: {len(AUDIT_LOG)}')\n", "print(f'成功执行: {executed}')\n", "print(f'被拒绝: {rejected}')\n", "print(f'拦截率: {rejected}/{len(AUDIT_LOG)} = {rejected/len(AUDIT_LOG)*100:.1f}%')" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 导出审计日志\n", "\n", "将审计日志保存为文件，作为交付物的一部分。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# 导出审计日志为 JSON\n", "audit_path = SANDBOX_ROOT / 'audit_log.json'\n", "with open(audit_path, 'w', encoding='utf-8') as f:\n", " json.dump(AUDIT_LOG, f, ensure_ascii=False, indent=2)\n", "\n", "# 同时导出为可读的文本格式\n", "audit_txt_path = SANDBOX_ROOT / 'audit.log'\n", "with open(audit_txt_path, 'w', encoding='utf-8') as f:\n", " for entry in AUDIT_LOG:\n", " ts = entry['timestamp']\n", " tool = entry['tool']\n", " final = entry['final']\n", " args_str = ', '.join(f'{k}={v}' for k, v in entry['arguments'].items())\n", " reason = ''\n", " if not entry['stages']['validation']['passed']:\n", " reason = f' (reason={entry[\"stages\"][\"validation\"][\"reason\"]})'\n", " f.write(f'[{ts}] PROPOSE tool={tool} {args_str} -> {final}{reason}\\n')\n", "\n", "print(f'审计日志已导出:')\n", "print(f' JSON: {audit_path}')\n", "print(f' 文本: {audit_txt_path}')\n", "print()\n", "print('文本格式预览:')\n", "print(audit_txt_path.read_text(encoding='utf-8'))" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 9. B档修改：修改一个变量\n", "\n", "从以下两项中选择**一项**修改，修改前先写预测。\n", "\n", "**选项A**：修改一条 Schema 限制\n", "- 例如：把 `create_note` 的内容长度限制从 500 改为 1000\n", "- 或者：把文件名规则改为也允许 `.txt` 后缀\n", "\n", "**选项B**：新增一个低风险工具 `append_checklist`\n", "- 功能：向已有笔记末尾追加一行待办事项\n", "- 要求：定义完整的 JSON Schema、权限和风险等级\n", "- 测试：正常调用 + 至少2种攻击测试\n", "\n", "**运行前先写预测**。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# 我选择：选项___\n", "# \n", "# 修改前预测：\n", "# 这个修改会扩大什么能力：\n", "# 这个修改会新增什么风险：\n", "# 需要增加什么测试：\n", "#" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 选项A：修改 Schema 限制\n", "\n", "取消下面的注释并修改。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# === 选项A：修改 Schema 限制 ===\n", "\n", "# 示例1：修改内容长度限制\n", "# original_max = tools['create_note']['parameters']['properties']['content']['maxLength']\n", "# tools['create_note']['parameters']['properties']['content']['maxLength'] = 1000 # 修改这个值\n", "# print(f'内容长度限制: {original_max} → 1000')\n", "\n", "# 示例2：修改文件名规则，允许 .txt\n", "# original_pattern = tools['create_note']['parameters']['properties']['filename']['pattern']\n", "# tools['create_note']['parameters']['properties']['filename']['pattern'] = r'^[a-zA-Z0-9_\\u4e00-\\u9fa5\\-]+\\.(md|txt)$'\n", "# print(f'文件名规则: {original_pattern}')\n", "# print(f' → {tools[\"create_note\"][\"parameters\"][\"properties\"][\"filename\"][\"pattern\"]}')\n", "\n", "# 修改后，重新测试正常用例和攻击用例\n", "# 观察：哪些之前被拒绝的调用现在可以通过了？\n", "# 新增了什么风险？" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 选项B：新增低风险工具\n", "\n", "取消下面的注释并修改。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# === 选项B：新增 append_checklist 工具 ===\n", "\n", "# 定义新工具的 Schema\n", "# new_tool = {\n", "# 'name': 'append_checklist',\n", "# 'description': '向已有笔记末尾追加一行待办事项',\n", "# 'parameters': {\n", "# 'type': 'object',\n", "# 'properties': {\n", "# 'filename': {\n", "# 'type': 'string',\n", "# 'description': '已有笔记的文件名',\n", "# 'pattern': '^[a-zA-Z0-9_\\\\u4e00-\\\\u9fa5\\\\-]+\\\\.md$'\n", "# },\n", "# 'item': {\n", "# 'type': 'string',\n", "# 'description': '要追加的待办事项',\n", "# 'maxLength': 100\n", "# }\n", "# },\n", "# 'required': ['filename', 'item'],\n", "# 'additionalProperties': False\n", "# },\n", "# 'permissions': {\n", "# 'filesystem': 'append',\n", "# 'scope': 'sandbox_only',\n", "# 'network': 'none'\n", "# },\n", "# 'risk_level': 'low'\n", "# }\n", "\n", "# 注册新工具\n", "# tools['append_checklist'] = new_tool\n", "\n", "# 实现执行逻辑（在 execute_tool 函数中添加）\n", "# elif tool_name == 'append_checklist':\n", "# filepath = SANDBOX_ROOT / arguments['filename']\n", "# if not filepath.exists():\n", "# return {'status': 'error', 'message': 'file_not_found'}\n", "# with open(filepath, 'a', encoding='utf-8') as f:\n", "# f.write('\\n- ' + arguments['item'])\n", "# return {'status': 'appended', 'file': arguments['filename'], 'item': arguments['item']}\n", "\n", "# 测试新工具\n", "# print('测试 append_checklist:')\n", "# r1 = process_tool_call('append_checklist', {'filename': '今日清单.md', 'item': '4. 复习攻击测试'})\n", "# print(f' 正常调用: {r1[\"log\"][\"final\"]}')\n", "# r2 = process_tool_call('append_checklist', {'filename': '../../etc/passwd', 'item': 'hacked'})\n", "# print(f' 路径穿越: {r2[\"log\"][\"final\"]}')\n", "# r3 = process_tool_call('append_checklist', {'filename': 'test.md', 'item': 'x'*200})\n", "# print(f' 超长内容: {r3[\"log\"][\"final\"]}')" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 修改后验证\n", "\n", "修改后重新运行正常用例和攻击用例，确认修改生效且没有引入新问题。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# 修改后验证\n", "# 取消注释并运行你需要的测试\n", "\n", "# 重新运行正常用例\n", "# print('修改后的正常用例测试:')\n", "# r = process_tool_call('create_note', {'filename': 'test_b.md', 'content': 'B档修改测试'})\n", "# print(f' 创建笔记: {r[\"log\"][\"final\"]}')\n", "\n", "# 重新运行攻击用例\n", "# print('修改后的攻击测试:')\n", "# r = process_tool_call('create_note', {'filename': '../../etc/passwd', 'content': 'hacked'})\n", "# print(f' 路径穿越: {r[\"log\"][\"final\"]} (应仍为 REJECTED)')\n", "\n", "# 我的观察：\n", "# 修改前后对比：\n", "# 新增风险分析：" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 10. 清理与验收\n", "\n", "清理沙盒目录，运行验收检查。" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# 清理沙盒\n", "shutil.rmtree(SANDBOX_ROOT, ignore_errors=True)\n", "print(f'沙盒已清理: {SANDBOX_ROOT}')\n", "print()\n", "\n", "# 验收清单\n", "checks = {\n", " '加载了工具 Schema': True,\n", " '运行了正常调用流程': True,\n", " '对比了\"模型声称\"与\"日志证明\"': True,\n", " '运行了攻击测试': True,\n", " '查看了审计日志': True,\n", " '写了攻击分析': False, # 检查第7节的分析单元格\n", " '完成了B档修改': False, # 检查第9节是否有修改\n", "}\n", "\n", "print('验收清单:')\n", "for check, status in checks.items():\n", " print(f' [{\"✓\" if status else \" \"}] {check}')\n", "\n", "print()\n", "print('请手动把 False 改为 True，确认你完成了对应项目。')\n", "print()\n", "print('关键概念检查:')\n", "print(' 1. 工具调用中，模型的角色是 ___，宿主程序的角色是 ___')\n", "print(' 2. JSON Schema 的作用是 ___')\n", "print(' 3. 路径穿越攻击的原理是 ___，防御方法是 ___')\n", "print(' 4. 审计日志的价值是 ___')\n", "print(' 5. \"模型声称已创建\"和\"日志证明已执行\"的区别是 ___')" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "name": "python", "version": "3.10.0" } }, "nbformat": 4, "nbformat_minor": 4 }