agent-browser

Warn

Audited by Gen Agent Trust Hub on Feb 19, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS] (MEDIUM): The skill requires installing the agent-browser package globally from npm. Since the author and organization are not in the trusted list, this introduces unverified code into the environment.
  • [REMOTE_CODE_EXECUTION] (MEDIUM): Executing agent-browser install triggers the download of Chromium binaries. This constitutes remote code execution from an unverified source.
  • [COMMAND_EXECUTION] (MEDIUM): The use of --with-deps during installation modifies OS-level packages. This typically requires elevated privileges and can lead to unauthorized system changes.
  • [PROMPT_INJECTION] (LOW): The skill is susceptible to indirect prompt injection as it ingests untrusted data from URLs. 1. Ingestion points: agent-browser open and snapshot operations in SKILL.md. 2. Boundary markers: Absent in instructions. 3. Capability inventory: Full browser control including element interaction (click, fill) and data extraction (snapshot, screenshot). 4. Sanitization: No evidence of input validation or escaping for web content is present in the instructions.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Feb 19, 2026, 05:19 PM