The Agent Skills Directory

[Indirect Prompt Injection] (MEDIUM): The skill accepts a --prompt argument that is passed directly to the Google Veo API.
Ingestion points: args.prompt in scripts/generate_video.py.
Boundary markers: Absent. The input is interpolated directly into the API request.
Capability inventory: File system write (generated_video.video.save), network access (Google API calls), and directory creation.
Sanitization: Absent. There is no filtering or validation of the prompt content before transmission.
[Arbitrary File Write] (MEDIUM): The --filename argument is converted to a Path object and used to create directories and save files without validation.
Evidence: output_path = Path(args.filename) followed by output_path.parent.mkdir(...) and generated_video.video.save(str(output_path)) in scripts/generate_video.py.
Risk: An attacker could provide a path like ../../.bashrc to overwrite sensitive user files if the agent executes the tool with broad permissions.
[External Downloads] (LOW): The script downloads the generated video file from Google's servers.
Evidence: client.files.download(file=generated_video.video) in scripts/generate_video.py.
Status: Target is a Google API, which is a trusted source, but the download is triggered by untrusted user input.

veo