agent-browser

Fail

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: HIGHPROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
  • Indirect Prompt Injection (HIGH): The skill is designed to ingest and process untrusted data from arbitrary web pages via agent-browser open, snapshot, and get text. It lacks explicit boundary markers or sanitization logic. Because the skill possesses high-privilege capabilities such as JavaScript execution (eval), form interaction (fill/click), and file uploads, an attacker-controlled website could potentially seize control of the agent's session.
  • Ingestion Points: agent-browser open <url>, agent-browser snapshot, agent-browser get text.
  • Capability Inventory: eval, cookies, storage local, upload, network route, fill, click.
  • Boundary Markers: None identified in SKILL.md.
  • Sanitization: None identified.
  • Data Exposure & Exfiltration (HIGH): Several commands provide direct access to sensitive session data which could be exfiltrated if the agent is compromised by a malicious site.
  • Evidence: agent-browser cookies (extracts session tokens), agent-browser storage local (accesses application data), and agent-browser state save auth.json (persists authentication state to disk).
  • Dynamic Execution (MEDIUM): The agent-browser eval command allows for the execution of arbitrary JavaScript within the context of the loaded web page. This can be used to bypass UI-based restrictions or programmatically extract data not visible in the accessibility tree.
  • Command Execution (LOW): The skill relies on the Bash(agent-browser:*) tool. While the usage examples appear standard for a browser automation CLI, the underlying binary's provenance is not established in the markdown file.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Feb 17, 2026, 04:40 AM