Model Deploy Skill

Use this skill when users request to deploy LLMs (Qwen, DeepSeek, etc.) on specified GPU servers and start the model service. This skill can Download models...

This skill appears to do what it says, but take these precautions before using it: - Ensure the agent host has deliberate SSH key-based access to target servers; do not provide private keys to unknown code. The script expects passwordless SSH. - Review and control the target server environment: the script will pip install packages, download large model files (disk/ bandwidth usage), and start a service listening on 0.0.0.0 — ensure firewalling and authentication as needed. - Confirm whether ModelScope models you will download are public or require credentials; this skill does not declare ModelScope credentials. - Validate the Miniconda location and conda availability on the target; the script expects $HOME/miniconda3 by default and will exit if not found. - Be cautious with PROXY env var values (they can redirect network traffic); set them only to trusted proxies. - Test on a non-production host first to verify behavior, port selection, GPU/memory usage, and to avoid accidental exposure of the model service.

Type: OpenClaw Skill Name: model-deploy Version: 1.0.0 The skill is classified as suspicious due to shell injection vulnerabilities in `scripts/deploy.sh`, where variables such as `${MODEL_PATH}` and `${MODEL_NAME}` are used unquoted in commands like `mkdir` and `vllm serve`. Furthermore, the instructions in `SKILL.md` direct the agent to perform remote execution via SSH using user-provided parameters (e.g., model name, organization) without explicit sanitization, which could be exploited to achieve Remote Code Execution (RCE) on the target GPU server.