The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted media content (audio/video) for transcription and dubbing. It includes explicit instructions to treat transcripts and lyrics as data rather than commands and provides evaluation cases (eval 13) specifically testing the agent's resistance to instructions embedded within transcribed text.
[COMMAND_EXECUTION]: Interacts with the ElevenLabs API through a bundled Python helper script (scripts/elevenlabs_media.py). The script is dependency-free, uses standard library HTTP clients, and strictly validates API paths and identifiers using regular expressions to prevent path traversal or injection attacks.
[EXTERNAL_DOWNLOADS]: Downloads are restricted to official ElevenLabs domains (api.elevenlabs.io, *.elevenlabs.io) and authorized Google Cloud Storage buckets (storage.googleapis.com). The helper script enforces HTTPS and rejects unrecognized origins or plain-text protocols.
[CREDENTIALS_UNSAFE]: The skill manages the ELEVENLABS_API_KEY but includes specific rules to prevent the key from being printed in logs, shown in chat, or leaked to third-party storage providers during downloads.
[DATA_EXPOSURE]: Implements local file security by creating output files and metadata receipts with restrictive permissions (0o600) and uses atomic file operations (exclusive creation and hard-linking) to prevent data races or unauthorized access.

elevenlabs-media