agent-browser

Fail

Audited by Gen Agent Trust Hub on Feb 15, 2026

Risk Level: HIGHEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [Indirect Prompt Injection] (HIGH): The skill is designed to process content from arbitrary web pages which can contain adversarial instructions meant to hijack the agent's behavior. 1. Ingestion points: agent-browser open, snapshot, and get text commands in SKILL.md. 2. Boundary markers: Absent. 3. Capability inventory: Interaction commands like click, fill, and state save allow a hijacked agent to perform sensitive actions. 4. Sanitization: Absent.
  • [External Downloads] (HIGH): The installation instructions require downloading a global npm package from an unverified source and subsequently downloading a binary (Chromium).
  • [Data Exposure & Exfiltration] (MEDIUM): The skill can save and load authentication states (auth.json), which contains sensitive session tokens. If an attacker can influence the agent via an untrusted page, they could potentially steal these tokens.
  • [Command Execution] (LOW): The skill executes various CLI commands to control the browser, which increases the impact of any potential injection attack.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Feb 15, 2026, 09:36 PM