agent-browser-3

Warn

Audited by Gen Agent Trust Hub on Mar 4, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The agent-browser eval command allows for the execution of arbitrary JavaScript code within the browser context, which can be used to manipulate page state or access internal browser data.
  • [COMMAND_EXECUTION]: The --executable-path global option enables the agent to specify and execute any local binary, potentially allowing for the execution of non-browser malicious files if the path is manipulated.
  • [DATA_EXFILTRATION]: The skill provides tools for reading and exporting sensitive session data. Commands such as agent-browser cookies, agent-browser storage, and agent-browser state save can be used to extract authentication tokens and session identifiers, which are then stored in local JSON files (e.g., auth.json, auth-state.json).
  • [PROMPT_INJECTION]: The skill exhibits a significant vulnerability surface for Indirect Prompt Injection (Category 8) due to its core function of processing untrusted web data.
  • Ingestion points: The agent ingests external, untrusted content through commands like agent-browser open, snapshot, get text, and get html across all provided templates and documentation.
  • Boundary markers: No boundary markers, delimiters, or explicit instructions to ignore embedded commands are present in the skill's logic or documentation.
  • Capability inventory: The skill possesses high-impact capabilities including arbitrary JavaScript execution (eval), form interaction (fill, click), and credential management (set credentials).
  • Sanitization: No sanitization, validation, or filtering of the retrieved web content is performed before it is added to the agent's context.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 4, 2026, 02:22 AM