agent-browser

Warn

Audited by Gen Agent Trust Hub on Jun 17, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONCREDENTIALS_UNSAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill is designed around the execution of the agent-browser CLI tool to perform web automation tasks. This involves executing shell commands with various arguments to control browser behavior.\n- [REMOTE_CODE_EXECUTION]: The eval command (detailed in references/commands.md) allows for the execution of arbitrary JavaScript within the browser context. The tool supports Base64 encoding and stdin to facilitate the execution of complex scripts, which presents a significant risk if the agent is directed to execute untrusted code.\n- [CREDENTIALS_UNSAFE]: Documentation and templates contain hardcoded example credentials that could be accidentally used or adapted unsafely: references/authentication.md contains the example password password123 and a hardcoded session token abc123xyz, while references/proxy-support.md includes examples with hardcoded proxy credentials username:password.\n- [DATA_EXFILTRATION]: The tool has the capability to read and save sensitive session data, including cookies, localStorage, and authentication states to local files. This data could be exfiltrated if the agent is compromised or if the state files are not properly secured.\n- [PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection as it processes untrusted content from the web.\n
  • Ingestion points: Web page content extracted via snapshot and get text (found in SKILL.md and references/snapshot-refs.md).\n
  • Boundary markers: The skill documentation (references/ADVANCED.md) recommends using --content-boundaries to mitigate this risk, but it is an opt-in feature and not enabled by default.\n
  • Capability inventory: The skill allows for full browser interaction, file system writes, and arbitrary network requests.\n
  • Sanitization: The tool does not appear to sanitize web content by default before presenting it to the agent.\n- [EXTERNAL_DOWNLOADS]: The documentation (references/ADVANCED.md) suggests installing external dependencies such as appium via npm for mobile testing functionality.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jun 17, 2026, 10:43 AM
Security Audit — agent-trust-hub — agent-browser