agent-browser

Pass

Audited by Gen Agent Trust Hub on Mar 1, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: The skill presents an indirect prompt injection surface because it is designed to interact with and process data from untrusted web pages.\n
  • Ingestion points: External content is ingested through the snapshot, get-text, get-html, and eval commands.\n
  • Boundary markers: There are no specific instructions or delimiters provided to help the agent isolate potentially malicious instructions embedded in web pages.\n
  • Capability inventory: The agent has the ability to navigate arbitrary URLs, write files (screenshots and state saves), and execute JavaScript.\n
  • Sanitization: No methods for sanitizing or validating external web content are described.\n- [COMMAND_EXECUTION]: The skill facilitates the execution of scripts and commands within the browser and the host sandbox.\n
  • Evidence: It includes a command for arbitrary JavaScript execution (eval) and troubleshooting instructions that involve using sudo via SSH (sudo chown -R 1000:1000 /home/user/.npm) to manage file permissions in the sandbox environment.\n- [EXTERNAL_DOWNLOADS]: The skill references the agent-browser package from Vercel Labs' official GitHub repository as its core automation engine.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 1, 2026, 03:43 PM