xhs-skill

Pass

Audited by Gen Agent Trust Hub on Jul 5, 2026

Risk Level: SAFEDATA_EXFILTRATIONPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [DATA_EXFILTRATION]: The scripts/review_publish_payload.mjs script reads files from paths provided in the media array of a user-controlled payload and sends their Base64-encoded content to an external AI provider. While it filters for specific image extensions (.png, .jpg, .jpeg, .webp), it lacks directory restriction or content verification, which could be exploited to read other sensitive files if they are accessible and renamed to match allowed extensions.
  • [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface in its content review mechanism. The scripts/review_publish_payload.mjs script builds a prompt for a sub-agent by directly interpolating user-provided text (title, body, tags) into the instructions. This lacks boundary markers or sanitization, potentially allowing an attacker to manipulate the reviewer AI into approving prohibited content.
  • Ingestion points: data/publish_payload.json (ingested by scripts/review_publish_payload.mjs)
  • Boundary markers: Absent; user text is appended directly to labels like payload: and body:.
  • Capability inventory: File system access (readFile) and network communication (fetch to AI provider) are active in the review script.
  • Sanitization: The skill performs heuristic policy checks for links and AI templates but does not escape or validate input against prompt injection payloads.
  • [COMMAND_EXECUTION]: The workflow relies on the execution of multiple shell commands via agent-browser-stealth and local Node.js utilities. This includes browser manipulation, cookie management, and image processing with the system sips tool. While these are necessary for the skill's function, they provide a broad surface for command injection if user inputs are not strictly controlled.
  • [EXTERNAL_DOWNLOADS]: The skill requires the installation of three third-party Node.js dependencies (jsqr, pngjs, and qrcode-terminal) from the NPM registry. These are standard packages for QR decoding and image handling but introduce external code into the execution environment.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 5, 2026, 09:05 AM
Security Audit — agent-trust-hub — xhs-skill