browser-use

Warn

Audited by Gen Agent Trust Hub on Oct 26, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The scripts/mcp-client.py script uses subprocess.Popen with shell=True to execute commands provided via the --stdio argument. This creates a risk of command injection if the agent is manipulated into passing malicious strings to this parameter.
  • [REMOTE_CODE_EXECUTION]: The scripts/start-server.sh script uses npx @playwright/mcp@latest to download and execute the Playwright MCP server from the NPM registry at runtime.
  • [DYNAMIC_EXECUTION]: The skill includes tools such as browser_evaluate and browser_run_code that allow for the execution of arbitrary JavaScript and Playwright code within the browser session, as detailed in references/playwright-tools.md.
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary function is to browse and interact with external websites, which presents a significant attack surface for indirect prompt injection.
  • Ingestion points: Untrusted data enters the agent's context through tools like browser_navigate and browser_snapshot in SKILL.md.
  • Boundary markers: None are implemented to distinguish between browser content and agent instructions.
  • Capability inventory: The skill possesses capabilities for shell execution (mcp-client.py), arbitrary script evaluation (browser_evaluate), and file system interaction (browser_file_upload).
  • Sanitization: There is no evidence of sanitization or filtering applied to content retrieved from the web.
  • [DATA_EXPOSURE]: The browser_file_upload tool defined in references/playwright-tools.md accepts absolute file paths, which could be exploited to exfiltrate sensitive files from the host system if the agent is improperly instructed.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 26, 2026, 09:38 AM