The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of subprocess calls to execute external binaries (ffmpeg, whisper) and orchestrate other local skills (dogpile, ops-nzbgeek, agent-inbox, scillm, memory). While it implements path sanitization in utils.py and uses list-based arguments to mitigate shell injection, the high volume of external tool orchestration increases the overall attack surface.\n- [COMMAND_EXECUTION]: In taxonomy.py, the skill utilizes importlib.util to dynamically load the common_taxonomy module from a computed relative path outside its directory. This dynamic loading pattern is a security concern as it could be exploited to load and execute unauthorized code if the directory structure is compromised.\n- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting and parsing untrusted subtitle data (.srt files).\n
Ingestion points: Untrusted data enters the agent context through scenes.py (parse_subtitle_file) and transcribe.py (resolve_subtitle_file).\n
Boundary markers: Absent; extracted subtitle text and tags are stored in JSON manifests without specific delimiters or instructions for the LLM to ignore embedded commands.\n
Capability inventory: The skill possesses significant capabilities, including media extraction via ffmpeg in extract.py, transcription via whisper in transcribe.py, and network operations for movie acquisition in radarr.py.\n
Sanitization: Absent; the skill does not perform sanitization or filtering of the text content extracted from subtitles before processing it.\n- [EXTERNAL_DOWNLOADS]: The skill interacts with remote services (NZBGeek, Radarr) and uses the subliminal library in subs.py to search for and download subtitle files from various third-party providers.

ingest-movie