agent-browser

Pass

Audited by Gen Agent Trust Hub on Mar 11, 2026

Risk Level: SAFE
Full Analysis
  • [PROMPT_INJECTION]: Analyzed the surface for indirect prompt injection via web content.
  • Ingestion points: The skill ingests untrusted data from web pages via the agent-browser snapshot -i command (referenced in SKILL.md and references/snapshot-refs.md).
  • Boundary markers: Present. The documentation explicitly recommends the use of AGENT_BROWSER_CONTENT_BOUNDARIES=1 to delimit web content and reduce injection risks.
  • Capability inventory: The skill has access to tools like Bash and Write (SKILL.md), and can perform browser-level actions like click, fill, and eval (references/commands.md).
  • Sanitization: Present. Security hardening options include AGENT_BROWSER_ALLOWED_DOMAINS for network restrictions and AGENT_BROWSER_ACTION_POLICY for fine-grained action control.
  • [COMMAND_EXECUTION]: The skill facilitates browser automation through a dedicated CLI and supports executing JavaScript within the browser context using eval --stdin. These features are fundamental to the skill's primary purpose and are accompanied by documented security guardrails.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 11, 2026, 01:20 PM