The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill script scripts/make_commercial.py accesses sensitive API credentials by reading from the local filesystem at ~/.runway/credentials and retrieving ELEVENLABS_API_KEY from environment variables. Accessing credential files directly from the user's home directory is a sensitive operation used here for authentication.
[COMMAND_EXECUTION]: The orchestrator and music generation scripts make extensive use of subprocess.run to execute external binaries including ffmpeg, magick (ImageMagick), and python3 to composite video, generate audio, and manage the project workflow.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to interact with ElevenLabs and Runway ML services. It sends text and image data to these platforms and downloads the resulting generated media assets to the local project directory.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection. Untrusted data provided in the brief.json configuration file (such as scene prompts, voiceover text, and headlines) is interpolated into shell commands for ffmpeg and magick.
Ingestion points: Scene fields (vo, prompt, headline, subhead) within brief.json and asset filenames.
Boundary markers: None used to isolate user-provided text within shell arguments.
Capability inventory: Subprocess execution of media tools and Python, and network access via the urllib library.
Sanitization: Minimal sanitization is performed (e.g., replacing newlines with spaces), which may not prevent injection of tool-specific control characters or escape sequences.

ai-commercial