The Agent Skills Directory

[DATA_EXFILTRATION]: The script scripts/confucius4_tts.py is capable of transmitting local file contents to a remote endpoint.
Evidence: The upload_reference_audio function in scripts/confucius4_tts.py reads the entire content of a file using reference_audio.read_bytes() and uploads it via a POST request to https://confucius4-tts.youdao.com/gradio/upload.
Risk: There are no checks to ensure the file provided via the --reference argument is an audio file. If an agent is manipulated by a malicious prompt into specifying a sensitive file path (e.g., ~/.ssh/id_rsa or .env), the script will read and exfiltrate that data to the remote server.
[COMMAND_EXECUTION]: The skill instructs the agent to execute a Python script that performs both network operations and filesystem writes.
Evidence: The script downloads synthesis results from the Gradio API and writes them to a local path using output.write_bytes(request_bytes(...)).
Risk: Without proper path validation or sanitization of the --out argument, the script could be used to overwrite existing local files or write data to unintended directories if the output path is maliciously provided.

voice-clone