sjht-data-annotation

通用数据标注处理工具。当用户提到需要数据标注、有标注任务、数据处理、数据集生成、 标注查看/编辑时使用此 skill。支持图像、视频、文本等多种数据类型，调用模型进行内容理解 和标注，生成结构化标注数据，提供 Web 查看编辑界面。 触发短语：「标注」「annotation」「数据集」「label」「tag da...

What to consider before installing/using this skill: - Functional fit: The skill matches a data-annotation workflow and includes a viewer and a small HTTP API. If you only need a local ad-hoc annotator, it is plausible to use. - Security issues to fix or review before use: - Restrict the API to the intended data directory: modify scripts/annotation-api.py so both listing (dir param) and save (results_file) are validated to reside under the configured DATA_DIR (use realpath checks). Right now GET allows listing arbitrary dirs and POST can write to arbitrary paths. - Do NOT chmod /root to 755. Instead place annotation data under a dedicated directory you control (e.g., /var/lib/annotation or a user home subdirectory) and run the service under a dedicated unprivileged user. - Avoid exposing the API publicly. If you proxy via nginx, require authentication, limit allowed origins, and avoid setting Access-Control-Allow-Origin: * for the API. Consider binding to a unix socket or 127.0.0.1 only and restricting nginx to authenticated locations. - Sanitize inputs in the POST save path and ensure files are not overwritten unintentionally. Prefer appending JSONL within DATA_DIR and disallow absolute paths in client-supplied file parameters. - Run the API and viewer in an isolated environment (container or dedicated VM) to limit impact if abused. - Operational cautions: - The SKILL.md suggests restarting nginx and modifying system config; perform these steps only if you understand the server and have backups. Prefer adding locations in an existing site config without changing root permissions. - The skill references external model APIs and examples that require API keys. Ensure you supply keys securely (environment variables, secret store) and do not embed them in client-side code. - If you want to proceed safely: review and patch annotation-api.py to enforce DATA_DIR bounds for both listing and saving; change default DATA_DIR to a non-root path; remove or rework instructions that require changing /root permissions; and run behind authentication/firewall. If you cannot audit/patch the code, treat the skill as risky and avoid exposing the service to other users or the internet.

Type: OpenClaw Skill Name: sjht-data-annotation Version: 1.0.0 The skill bundle provides a data annotation workflow with a web-based viewer, but it contains significant security vulnerabilities and risky instructions. The `scripts/annotation-api.py` script allows writing data to arbitrary file paths via the `file` parameter in POST requests, lacking the path-traversal protections found in its file-reading logic. Additionally, `SKILL.md` instructs the agent to perform high-risk system modifications, such as changing the `/root` directory permissions to `755` and altering the default Nginx configuration. While these actions are contextually related to deploying the annotation viewer, they introduce severe security flaws without evidence of direct malicious intent.