The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches and transcribes audio from external podcast episodes (Spotify, RSS feeds, or direct MP3 URLs). A malicious actor could embed verbal instructions in the podcast audio that, once transcribed, might influence the LLM's behavior during the summarization phase.
Ingestion points: scripts/summarize_podcast.py fetches content from arbitrary external URLs provided by the user or found in RSS feeds.
Boundary markers: The transcription text is interpolated directly into the prompt template (Transcript:\n{text[:50000]}) without robust isolation markers or instructions to ignore embedded commands.
Capability inventory: The skill possesses the ability to execute shell commands (whisper), make network requests (requests), and write to the local file system.
Sanitization: There is no sanitization or filtering of the transcribed text before it is sent to the LLM for summarization.
[COMMAND_EXECUTION]: The skill uses subprocess.run to execute the whisper CLI tool for speech-to-text processing. While the implementation follows best practices by passing arguments as a list rather than a shell string (which mitigates shell injection), it remains an execution of a local system binary.

podcast-summarizer