clipchamp-video
Pass
Audited by Gen Agent Trust Hub on Aug 30, 2026
Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [Sensitive Data Access]: The skill includes automation to clone Microsoft Edge user profiles, including cookies and local storage, to a local working directory (
~/.copilot/video-work/edge-tenant). This is a design choice to maintain authentication for automated screen recording of protected applications like Copilot Studio. - [System Command Execution]: The skill makes extensive use of PowerShell scripts and Node.js subprocesses to manage window positioning, terminate browser instances, and execute
ffmpegfor screen capture and video encoding. - [External Dependency Management]: The skill requires and installs several standard tools via well-known package managers (
npm install -g playwright,pip install edge-tts, andwinget install --id Gyan.FFmpeg). These are established utilities used for the skill's core video production features. - [Indirect Prompt Injection Surface]: The recorder script monitors agent responses to detect completion (e.g., by checking for feedback buttons) and trigger narration beats. This creates a surface where the agent's output influences the automation timing.
Ingestion points:
- Agent response text and UI elements are monitored via Playwright locators in
scripts/recorder.js.
Boundary markers:
- None observed for agent response monitoring.
Capability inventory:
- File system read/write (cloning profiles, writing logs, saving video).
- Process management (killing Edge and Node processes).
- System command execution (PowerShell scripts, ffmpeg).
- Network operations (Playwright browser navigation, TTS generation).
Sanitization:
- The skill uses standard Playwright selectors for interaction and does not appear to execute instructions embedded within the agent's text responses.
Audit Metadata