ui-video-builder
Warn
Audited by Gen Agent Trust Hub on Nov 18, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: Multiple scripts, including
scripts/check.mjs,scripts/repeat.mjs, andreferences/samples/screen/render.mjs, utilize dynamicimport()calls to load adapter and scene scripts from filesystem paths determined at runtime. This design allows for a plugin-like architecture but results in the execution of local JavaScript files specified via command-line arguments or configuration. - [DATA_EXFILTRATION]: The
references/samples/screen/record.mjsscript records all HTTP and WebSocket communication between the browser and the application's backend. It captures full request and response bodies, saving them as base64-encoded strings in local JSON files. Although documentation advises on masking secrets and using.gitignore, the absence of automated sanitization in the recording script poses a risk of accidental exposure of session tokens, API keys, or personal data. - [REMOTE_CODE_EXECUTION]: The narration generation script (
references/samples/screen/narrate.mjs) interacts with the ElevenLabs text-to-speech API. This is a legitimate integration with a well-known service for the skill's intended purpose, utilizing an API key provided via environment variables. - [EXTERNAL_DOWNLOADS]: The skill requires installing the Playwright framework and associated browser binaries (Chromium). These are industry-standard developer tools provided by a trusted organization.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from network recordings and scene scripts. Evidence shows ingestion in
record.mjsandrender.mjswith no explicit boundary markers or sanitization logic. The skill possesses capabilities for file writing, network requests, and subprocess execution (via FFmpeg and ffprobe), creating a surface where malicious application data could theoretically influence the agent's behavior during video refinement or debugging tasks.
Audit Metadata