The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external command-line tools ffprobe and mediainfo to perform deep media analysis. This is handled in scripts/lib/cmd.py using subprocess.run with list-based arguments, which is a secure practice that prevents shell injection from malicious file paths or URLs.
[EXTERNAL_DOWNLOADS]: In scripts/hls_check.py, the skill fetches remote HLS playlists and video segments using the urllib.request library. While this is the intended functionality for HLS diagnostics, it involves downloading up to 24 MB of data per run from user-provided URLs. The script implements a 15-second timeout and a per-segment byte cap (8 MB) to prevent resource exhaustion.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface because it parses untrusted media files and HLS manifests. Malicious metadata embedded in these files (such as titles, comments, or custom HLS tags) could be extracted into the JSON reports generated by the scripts. If the AI agent processes this extracted metadata without sanitization, it could interpret injected text as instructions.
Ingestion points: User-provided media files and URLs are parsed by ffprobe (via scripts/media_analyze.py) and HLS manifests are parsed by scripts/hls_check.py.
Boundary markers: The output is structured as JSON, which provides structural delimitation, but the content within fields like summary or probe_summary is interpreted by the LLM.
Capability inventory: The skill has the capability to read local files, fetch network resources, and execute pre-installed media tools.
Sanitization: There is no evidence of explicit filtering or sanitization of extracted string metadata (e.g., codec profiles, format tags) before it is passed to the agent context.

alibabacloud-media-diagnostics