browser-automation

Warn

Audited by Gen Agent Trust Hub on Nov 7, 2026

Risk Level: MEDIUMCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [CREDENTIALS_UNSAFE]: The skill includes scripts for automated extraction of sensitive identity tokens. Specifically, extension-token-from-profile.ts accesses internal Chrome storage (LevelDB) to retrieve authentication tokens for the Playwright extension.
  • [CREDENTIALS_UNSAFE]: The fill-secret.mjs utility programmatically retrieves secrets from the macOS system Keychain to automate password entry in browser forms.
  • [INDIRECT_PROMPT_INJECTION]: As a browser automation tool, the skill provides a substantial surface for indirect prompt injection. Malicious websites could attempt to control the agent via instructions embedded in page content.
  • Ingestion points: Browser page content accessed via Playwright/CDP.
  • Boundary markers: The doctrine recommends profile isolation but the skill lacks explicit boundary markers for untrusted content.
  • Capability inventory: Keychain access, system script execution (bash/node/bun), and full browser control.
  • Sanitization: Operational rules suggest using form.requestSubmit() and verifying state changes independently to prevent bot detection and command misinterpretation.
  • [COMMAND_EXECUTION]: The skill executes several bundled scripts (chrome-profile.sh, chrome-tabs.ts, etc.) using bash, node, and bun to control system behavior and interface with Chrome via AppleScript.
  • [EXTERNAL_DOWNLOADS]: The skill uses npx -y @playwright/mcp@latest to install and run the browser automation server from the npm registry.
  • [DATA_EXFILTRATION]: The tabs list feature allows the agent to see all open tab URLs and titles across all Chrome profiles, potentially exposing sensitive navigation history or active session metadata.
  • [DYNAMIC_EXECUTION]: The skill supports dynamic JavaScript execution within the browser using browser_run_code_unsafe and drives tool sequences via the mcp-drive.ts script.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 7, 2026, 02:12 AM