The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The process_asset.py script employs the trust_remote_code=True parameter when loading the BiRefNet model from Hugging Face. This setting allows the repository to execute arbitrary Python code on the local machine during the model loading process.
[COMMAND_EXECUTION]: The skill's workflow instructs the agent to run shell commands that incorporate variables like {target_size} (from user input) and {downloaded_file} (from the file system). This creates a command injection surface if the agent executes these strings without proper sanitization or escaping.
[EXTERNAL_DOWNLOADS]: The skill downloads a large ML model (~1GB) from Hugging Face when the --ml flag is used for background removal.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted data (user arguments and external filenames) to perform sensitive operations.
Ingestion points: User-provided CLI arguments (--size) and the names of files found in the ~/Downloads folder.
Boundary markers: No delimiters or security warnings are used when interpolating these values into shell commands.
Capability inventory: The skill uses the shell to list files, move files, create directories, and run Python scripts.
Sanitization: There is no evidence of input validation or escaping logic in the provided Python script or the workflow instructions.

2d-pixel-asset