vercel-agent-browser

Warn

Audited by Gen Agent Trust Hub on May 16, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill instructs users to install the agent-browser package from global registries (NPM, Cargo, Homebrew). It also utilizes an agent-browser install command which downloads external binaries (Chrome for Testing) at runtime.
  • [COMMAND_EXECUTION]: Provides high-privilege automation capabilities including agent-browser eval for arbitrary JavaScript execution and agent-browser batch for multi-step command sequences.
  • [DATA_EXFILTRATION]: Demonstrates patterns for accessing sensitive local and environment data, such as uploading local files via agent-browser upload, retrieving browser cookies/storage, and interpolating environment variables like ${PASSWORD} into CLI commands.
  • [PROMPT_INJECTION]: Vulnerable to indirect prompt injection by design, as it retrieves and processes untrusted content from arbitrary websites.
  • Ingestion points: agent-browser snapshot, agent-browser get text, and agent-browser eval (reading DOM content) in SKILL.md.
  • Boundary markers: Absent; the skill does not define delimiters or provide instructions to ignore embedded commands in retrieved web data.
  • Capability inventory: Includes file system access (upload), network interception (network route), and JavaScript evaluation (eval).
  • Sanitization: No sanitization or validation of external content is mentioned before it is processed by the agent.
  • [METADATA_POISONING]: The skill uses the name 'vercel-agent-browser', which may lead users to believe it is an official Vercel product, whereas the overview identifies the author as 'ara.so' (Aradotso).
Audit Metadata
Risk Level
MEDIUM
Analyzed
May 16, 2026, 03:27 PM
Security Audit — agent-trust-hub — vercel-agent-browser