browser-work

Pass

Audited by Gen Agent Trust Hub on May 20, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: Executes shell commands to initialize session directories in ~/.hoyeon/, generate session identifiers using openssl, and manage files using cat and redirection.
  • [EXTERNAL_DOWNLOADS]: Utilizes npx to download and execute the @team-attention/chromux package if it is not found in the system path. This package is a resource associated with the skill author.
  • [PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by ingesting untrusted data from external websites and using it to influence sub-agent behavior.
  • Ingestion points: Captures website accessibility trees and text via chromux snapshot in SKILL.md (Step 2).
  • Boundary markers: Site content is stored in a guide file and then interpolated directly into a sub-agent prompt without explicit boundary delimiters or instructions to ignore embedded commands.
  • Capability inventory: The browser-explorer sub-agent has capabilities to click, fill forms, type, and navigate pages.
  • Sanitization: There is no visible sanitization or filtering of the ingested web content before it is provided to the sub-agent.
Audit Metadata
Risk Level
SAFE
Analyzed
May 20, 2026, 01:29 PM
Security Audit — agent-trust-hub — browser-work