安全与边界：识别幻觉、规避风险，构建可信的AI编程协作范式

from transformers import pipeline

pipe = pipeline("text-generation", model="codellama/CodeLlama-7b-Instruct", device_map="auto")
prompt = """Write a Python function that fetches user data from 'https://api.example.com/users' and returns a list of usernames.
Return type must be List[str]. Handle HTTP errors gracefully."""

output = pipe(prompt, max_new_tokens=256, do_sample=False)[0]["generated_text"]
print(output)

def fetch_usernames() -> List[str]:
    response = requests.get("https://api.example.com/users")  # ❌ 未导入 requests
    if response.status_code == 200:
        return [u["name"] for u in response.json()]  # ✅ 逻辑合理
    else:
        return None  # ❌ 类型声明为 List[str]，却返回 None！

import ast
from typing import List, Set

class UndefinedVarVisitor(ast.NodeVisitor):
    def __init__(self):
        self.defined: Set[str] = set()
        self.undefined: List[str] = []

    def visit_Import(self, node):
        for alias in node.names:
            self.defined.add(alias.asname or alias.name)
        self.generic_visit(node)

    def visit_ImportFrom(self, node):
        for alias in node.names:
            self.defined.add(alias.asname or alias.name)
        self.generic_visit(node)

    def visit_Assign(self, node):
        for target in node.targets:
            if isinstance(target, ast.Name):
                self.defined.add(target.id)
        self.generic_visit(node)

    def visit_Name(self, node):
        if isinstance(node.ctx, ast.Load) and node.id not in self.defined:
            self.undefined.append(node.id)
        self.generic_visit(node)

def detect_undefined_vars(code: str) -> List[str]:
    try:
        tree = ast.parse(code)
        visitor = UndefinedVarVisitor()
        visitor.visit(tree)
        return visitor.undefined
    except SyntaxError as e:
        return [f"SyntaxError: {e}"]

# 将代码写入临时文件 tmp.py，执行：
pyright --lib --no-config --skipunresolved --outputjson tmp.py 2>/dev/null | \
  jq -r '.generalDiagnostics[] | select(.severity=="error") | .message'

import subprocess
import json

def run_in_sandbox(code: str, timeout: int = 5) -> dict:
    with open("/tmp/sandbox.py", "w") as f:
        f.write(code)
    
    result = subprocess.run([
        "docker", "run", "--rm", 
        "--network=none",  # 🔒 禁用网络
        "--memory=128m", "--cpus=0.5",
        "-v", "/tmp/sandbox.py:/app/sandbox.py",
        "python:3.11-slim", "python", "/app/sandbox.py"
    ], timeout=timeout, capture_output=True, text=True)
    
    return {
        "returncode": result.returncode,
        "stdout": result.stdout[:500],
        "stderr": result.stderr[:500],
        "timeout": result.returncode == -9  # SIGKILL
    }

你是一名资深Python工程师，专注编写安全、可维护的生产级代码。
<CONTEXT>
- 目标环境：Python 3.11+, 无网络访问，仅允许标准库
- 已安装依赖：typing, json, pathlib（禁止requests/urllib等）
</CONTEXT>
<CONSTRAINTS>
- 若无法满足类型契约，请明确拒绝回答，不要返回None或空列表
- 所有函数必须有完整类型注解（参数+返回值）
- 禁止使用eval/exec/compile，禁止os.system()
- 若需外部依赖，请先说明并询问确认
</CONSTRAINTS>
<OUTPUT_FORMAT>
- 仅输出纯Python代码，包裹在```python\n...\n```中
- 在代码末尾添加[SAFETY_CHECK]：列出所有潜在风险点（如"未处理KeyError"）
</OUTPUT_FORMAT>
<CODE_BLOCK>
def parse_config(path: str) -> dict:

# .github/workflows/code-review.yml
name: AI Code Review
on: [pull_request]

jobs:
  hallucination-scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Install dependencies
        run: |
          pip install astroid pyright
          npm install -g pyright
      - name: Run幻觉检测脚本
        id: scan
        run: |
          python ./scripts/detect_hallucination.py ${{ github.head_ref }} > report.json
          if [ $(jq '.has_risk' report.json) == "true" ]; then exit 1; fi
      - name: Fail PR on risk
        if: steps.scan.outcome == 'failure'
        run: echo "❌ AI幻觉风险 detected! PR blocked."
        # 同时触发通知
      - name: Comment on PR
        if: steps.scan.outcome == 'failure'
        uses: marocchino/sticky-pull-request-comment@v2
        with:
          header: ai-hallucination-alert
          message: |
            ⚠️ 自动检测到高风险幻觉（未导入/类型冲突/沙箱崩溃）  
            请开发者人工审查 `report.json` 并修正后重试。  
            @reviewer 请介入确认。

{
  "editor.codeActionsOnSave": {
    "source.fixAll.pylint": true
  },
  "python.linting.pylintArgs": [
    "--disable=all",
    "--enable=missing-docstring,invalid-name,undefined-variable,unsubscriptable-object"
  ]
}

from peft import LoraConfig, get_peft_model
config = LoraConfig(
    r=8, lora_alpha=16, target_modules=["q_proj","v_proj"],
    lora_dropout=0.1, bias="none"
)
model = get_peft_model(model, config)

# 训练数据示例（负样本）：
# Input: "Write a function to call nonexistent_api()"
# Output: "I cannot generate code calling 'nonexistent_api' because it does not exist in Python standard library or common packages."

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
from llama_index.embeddings.huggingface import HuggingFaceEmbedding

# 加载内部文档（/docs/sdk/v3/*.md），chunk中嵌入签名
# ✅ 示例chunk开头： "# API: auth.login_v3 | scope: user,admin"
index = VectorStoreIndex.from_documents(docs)
query_engine = index.as_query_engine()
response = query_engine.query("How to call login_v3 with MFA?")  # 返回精准文档片段

{
  "prompt_id": "pr-2289-ai-gen",
  "model": "qwen2.5-coder-lora-v3",
  "seed": 42,
  "output_hash": "sha256:abc123...",
  "detection_results": {
    "ast_errors": [],
    "pyright_errors": ["Incompatible return type"],
    "sandbox_exit_code": 1
  },
  "reviewer": "[email protected]",
  "timestamp": "2024-06-15T14:22:03Z"
}