record-ui-walkthrough
Pass
Audited by Gen Agent Trust Hub on Oct 10, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes
subprocess.runacross several scripts (build-video.py,check-kit.py,validate-video.py) to callffmpeg,ffprobe, andplaywright-cli. These calls are properly parameterized using lists to avoid shell injection and target standard system or local skill binaries. - [EXTERNAL_DOWNLOADS]: The skill's setup documentation (
references/setup.md) instructs the user to download Text-to-Speech model files from a third-party GitHub repository (thewh1teagle/kokoro-onnx). The skill mitigates potential supply-chain risks by verifying these files against hardcoded SHA-256 hashes inscripts/preview_common.pybefore execution. - [DYNAMIC_EXECUTION]: The skill employs
playwright-clito execute dynamically generated JavaScript code (hero.js,preflight.js) within a browser environment for automation purposes. Additionally,scripts/build-video-with-voice.pyuses Python'simportlib.utilto load rendering logic fromscripts/build-video.pyat runtime. - [INDIRECT_PROMPT_INJECTION]: The skill processes agent-provided descriptions and UI labels for video captions, creating an attack surface for subtitle manipulation.
- Ingestion points: Untrusted data enters via
shot-list.jsonandworkflow.json(generated from agent-defined hero scripts). - Boundary markers: The skill enforces strict length constraints (maximum 18 words) and prohibits newlines in descriptions.
- Capability inventory: Captions are rendered into video frames using
ffmpeg's libass filter. - Sanitization: The
ass_textfunction inscripts/build-video.pyescapes backslashes and curly braces using the\u2060(Word Joiner) character to prevent the injection of malicious formatting tags into the ASS subtitle files.
Audit Metadata