agent-browser

Pass

Audited by Gen Agent Trust Hub on Mar 1, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection via the web content it processes.
  • Ingestion points: The agent reads untrusted data from the internet using commands such as agent-browser open, agent-browser snapshot, and agent-browser get text (SKILL.md).
  • Boundary markers: No explicit delimiters or instructions are provided to the agent to treat external web content as data rather than instructions.
  • Capability inventory: The agent-browser CLI includes capabilities for network navigation, form interaction, and local file system writes (e.g., screenshot and state save).
  • Sanitization: No sanitization or content validation is performed on the data retrieved from web pages.
  • [COMMAND_EXECUTION]: The skill utilizes a custom CLI tool, agent-browser, to automate web interactions. While this is the intended functionality, it grants the agent a high degree of control over a browser environment which could be abused if the agent is successfully injected.
  • [DATA_EXFILTRATION]: The skill includes the ability to save sensitive browser session data, including cookies and authentication tokens, to the local file system using the agent-browser state save command. While a legitimate feature for automation, it represents a data exposure risk if not handled carefully.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 1, 2026, 12:37 AM