video-content-analysis
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses PowerShell to execute local scripts and manage a Python virtual environment. This is the primary mechanism for orchestration.\n- [DYNAMIC_EXECUTION]: The skill generates Python code snippets at runtime to invoke processing functions, interpolating variables such as file paths and configuration flags. It also patches
torch.loadbehavior for compatibility with specific model checkpoints.\n- [EXTERNAL_DOWNLOADS]: The installer fetches FFmpeg, Python runtimes, and AI models from official repositories including the Microsoft Store (winget), Python.org, and HuggingFace.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests external content from video transcripts and caption files (.vtt) and displays synthesized summaries in the agent's conversation window, creating a potential surface for indirect instruction injection. Mandatory Evidence Chain:\n - Ingestion points: Reads transcription results from Whisper/pyannote and external .vtt files (Step 4, SKILL.md).\n
- Boundary markers: None explicitly mentioned in the summary display instruction.\n
- Capability inventory: Includes file system writes, subprocess calls, and network access (orchestrator.py, privacy.py).\n
- Sanitization: Not explicitly implemented in the summary display step.\n- [SAFE]: Credentials (HuggingFace API tokens) are managed securely through the Windows Credential Manager (
cmdkey) rather than being hardcoded or stored in plain text.\n- [SAFE]: Telemetry and data processing are performed locally or within the user's Microsoft 365 environment, with explicit gates for sensitive content and consent status.
Audit Metadata