The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/matte_rvm.py performs remote code execution by calling torch.hub.load with trust_repo=True on the PeterL1n/RobustVideoMatting repository. This repository is not a recognized trusted vendor, and trust_repo=True explicitly permits the execution of arbitrary Python code defined in the remote repository's hubconf.py file.
[EXTERNAL_DOWNLOADS]: The skill relies on fetching external resources at runtime, including the HyperFrames CLI via npx and machine learning models for video matting and face detection. Several of these sources are third-party and fall outside the scope of trusted organizational vendors.
[DATA_EXFILTRATION]: The template/build.py script is vulnerable to path traversal. It uses paths supplied in the beats.json configuration file to copy files from the local filesystem to an output directory (using shutil.copyfile) without validation or sanitization. An attacker providing a malicious configuration could trick the skill into exposing sensitive files, such as environment variables or private keys, by using relative paths (e.g., ../../.env).
[COMMAND_EXECUTION]: Multiple utility scripts (build.py, common.py, matte_rvm.py, etc.) execute external binaries such as ffmpeg and ffprobe through the subprocess module. These executions are parameterized by values and file paths extracted from user-supplied JSON files, creating an injection surface if the inputs are maliciously crafted.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from reel.json, beats.json, and captions.json to drive its build and render pipeline. This ingestion surface allows for indirect injection that influences file operations, command-line arguments, and the behavior of the QA validation scripts.

interview-reel-motion