The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions direct the user to clone a codebase from an untrusted GitHub repository (ap229997/hands). While the metadata identifies the author as K-Dense Inc., the repository is hosted under an unverified personal account.
[REMOTE_CODE_EXECUTION]: The setup process includes executing a shell script (bash scripts/download_models.sh) and running a Python script (python demo.py) from the cloned repository. Since the source repository is not from a trusted vendor, this behavior represents unverifiable remote code execution.
[COMMAND_EXECUTION]: The skill requires several high-privilege command-line operations for installation and execution, including git clone, pip install -r requirements.txt, and bash script execution. These commands execute code from an external source without prior validation.
[INDIRECT_PROMPT_INJECTION]: The skill processes external video and image data which serves as a potential attack surface.
Ingestion points: Reads video files (egocentric_video.mp4) and images (frame.jpg) using cv2.VideoCapture and cv2.imread as described in SKILL.md.
Boundary markers: No boundary markers or instructions to ignore embedded content are present.
Capability inventory: The skill uses np.save, json.dump, and cv2.VideoWriter to write files to the local system.
Sanitization: No sanitization or validation of the input media files is mentioned.

hands-3d-pose