agent-browser

Pass

Audited by Gen Agent Trust Hub on Sep 23, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process content from arbitrary web pages, which could contain malicious instructions designed to influence the agent's behavior.
  • Ingestion points: The agent reads external data using agent-browser snapshot, agent-browser get text, and agent-browser get html as documented in SKILL.md.
  • Boundary markers: No specific boundary markers or instruction-following constraints are defined for the data retrieved from web pages.
  • Capability inventory: The skill possesses extensive capabilities including file uploads, network routing, session state persistence, and JavaScript execution.
  • Sanitization: The skill does not implement explicit sanitization of the content retrieved from the browser before presenting it to the agent.
  • [DYNAMIC_EXECUTION]: The tool provides direct methods to execute arbitrary JavaScript and conditional logic within the browser context.
  • Evidence: The agent-browser eval command allows execution of arbitrary strings as JavaScript, and agent-browser wait --fn allows execution of JavaScript conditions.
  • [DATA_EXFILTRATION]: Several features allow for the access and potential transmission of local or sensitive data.
  • Evidence: The agent-browser open command explicitly supports the file:// protocol, which can be used to read local system files into the browser context.
  • Evidence: The agent-browser upload command allows local files to be uploaded to web forms, facilitating potential data exfiltration.
  • Evidence: Commands like agent-browser cookies and agent-browser storage can extract sensitive session tokens, which the skill documents can be saved to local files via agent-browser state save.
  • [COMMAND_EXECUTION]: The skill facilitates the execution of the agent-browser CLI and allows configuration of the execution environment.
  • Evidence: The --executable-path option allows the agent to specify an arbitrary binary to be run as the browser executable, potentially leading to unauthorized command execution if the path is manipulated.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 23, 2026, 01:39 AM