ucloud-deepseek-ocr

OCR text recognition using DeepSeek-OCR model. Use when user asks for OCR, text recognition, image text extraction, screenshot recognition, or converting ima...

This script mostly does what the skill claims (encode an image and call a remote OCR model) but has a few red flags: 1) The SKILL.md path ({baseDir}/scripts/ocr.sh) does not match the provided file (script/ocr.sh) — verify the correct path before running. 2) The script sources ~/.openclaw-env if present; avoid running it if that file might contain untrusted commands or secrets. Consider opening that file to inspect its contents, or run the script in a controlled environment (container or sandbox). 3) Confirm you trust the API host (https://api.modelverse.cn by default) before providing DEEPSEEK_OCR_API_KEY. 4) If you want the key stored in ~/.openclaw/openclaw.json as documented, understand the script will not read that file unless you modify it. If you are not comfortable with these issues, request a corrected SKILL.md and/or a version of the script that reads a specified config file (without sourcing) or accepts the API key via an explicit argument or a securely-loaded env var.

Type: OpenClaw Skill Name: ucloud-deepseek-ocr Version: 1.0.0 The skill is classified as suspicious due to multiple shell injection vulnerabilities in `script/ocr.sh`. The `API_URL` variable is used unquoted in the `curl` command, allowing for arbitrary `curl` argument injection if an attacker controls the `DEEPSEEK_OCR_API_URL` environment variable. Additionally, the `IMAGE_PATH` variable is unquoted when passed to `base64` and in file existence checks, posing further shell injection risks. The skill also presents a prompt injection vulnerability against the DeepSeek-OCR model via the `output_format` parameter, allowing an attacker to manipulate the AI model's instructions.