html-export
Pass
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes HTML content from both local files and remote URLs. This ingestion of untrusted data represents a potential attack surface where malicious content could attempt to influence the agent via the exported media or report files.
- Ingestion points: HTML input files in
scripts/export.cjsand URLs inscripts/capture-sites.cjs(viatargets.json). - Boundary markers: Structured JSON reports are used for metadata. Output media (PNG, PDF, MP4, GIF) are non-textual.
- Capability inventory: Uses Playwright for rendering, ffmpeg for media encoding, and standard Node.js filesystem modules for writing outputs.
- Sanitization:
scripts/capture-sites.cjsimplements strict sanitization, truncation, and control-character removal for all page-derived metadata (e.g., titles and URLs) included in thecapture-manifest.json. - [COMMAND_EXECUTION]: The skill invokes external binaries such as
ffmpeg,ffprobe, and Chromium. All process executions utilize thechild_process.spawnAPI withshell: false, passing arguments as arrays to prevent shell command injection. Overrides for binary paths are required to be absolute and, on Windows, must use.exeor.comextensions to avoid the execution of potentially vulnerable batch wrappers. - [DATA_EXFILTRATION]: The skill provides an
--offlinemode for rendering local HTML files, which mitigates network-based data exfiltration. It blocks all non-local network traffic, including WebSockets and WebTransport, using a dead proxy and Playwright context routing. It also specifically prevents exfiltration via UNC/SMB paths by restricting allowedfile:URLs to the local host. While the renderer can still access local files as required for HTML processing (potentially including them in the media output), these protections effectively secure the execution environment against unauthorized remote transmission.
Audit Metadata