agent-browser

Warn

Audited by Gen Agent Trust Hub on Nov 3, 2026

Risk Level: MEDIUMDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill provides instructions for exposing local content to the internet to facilitate processing by remote vision models.
  • Evidence: The documentation suggests running a local Python HTTP server (python3 -m http.server) and exposing it via cloudflared tunnel to a public trycloudflare URL.
  • Risk: Creating a public tunnel to a local filesystem directory allows external actors to access those files if the URL is discovered or the server is misconfigured.
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary function involves scraping and interpreting content from external, untrusted websites.
  • Ingestion points: Commands such as agent-browser read, snapshot, and a11y fetch data directly from user-provided or agent-discovered URLs.
  • Boundary markers: There are no instructions for the agent to use specific delimiters or to ignore potential instructions embedded in the fetched markdown or accessibility trees.
  • Capability inventory: The skill possesses the ability to execute shell commands, manage files, and perform network requests, which could be abused if the agent follows malicious instructions from a scraped page.
  • Sanitization: The instructions do not describe any sanitization or validation steps for the data retrieved from external URLs before it is processed by the LLM.
  • [COMMAND_EXECUTION]: The skill relies on complex shell command sequences for its core functionality.
  • Evidence: The file includes command blocks for brew, agent-browser, ffmpeg, and cloudflared.
  • Context: Automated execution of these tools on untrusted data fetched from the internet increases the overall risk profile of the skill.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 3, 2026, 10:26 AM