app-preview-video
Pass
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The build script (scripts/build.mjs) reads a user-provided JavaScript file defined in the plan.json configuration and embeds it directly into an HTML overlay for execution within a headless browser. It also uses the Chrome DevTools Protocol to evaluate JavaScript expressions at runtime to manage the video capture sequence. Furthermore, the script uses the eval() function to parse frame rate strings returned by ffprobe.
- [COMMAND_EXECUTION]: The skill relies on spawning external processes to perform its primary tasks. It invokes ffmpeg and ffprobe for media processing, and agent-browser to orchestrate the frame-by-frame capture of the HTML overlay.
- [INDIRECT_PROMPT_INJECTION]: The skill provides an attack surface by ingesting data from the local repository (such as plan.json and media clips) and executing code from local scripts. A malicious repository could provide a crafted script that performs unauthorized actions when the agent attempts to build a video. Ingestion points: plan.json and the script file specified in the mock field (e.g., screens.js). Boundary markers: None; local scripts are executed verbatim in the browser context. Capability inventory: File system read/write, external command execution via node:child_process, and browser control via CDP. Sanitization: Basic HTML escaping is applied to text fields, but the mockup logic is treated as executable code.
Audit Metadata