The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides the agent-browser eval command, which allows for JavaScript execution within the browser. This is a standard feature for automation tasks and is documented with safe usage examples.
[DATA_EXFILTRATION]: Features for saving and loading session state (agent-browser state save/load) are included to support authenticated sessions. The documentation correctly identifies the sensitivity of these files and provides guidance on secure management.
[PROMPT_INJECTION]: The skill facilitates the ingestion of external web content, which presents a surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context via commands like agent-browser open and snapshot in SKILL.md and various templates.
Boundary markers: No specific delimiters are mandated for separating retrieved web content from agent instructions.
Capability inventory: The CLI tool allows for file access, network operations, and code execution within the browser environment.
Sanitization: Content retrieval is direct, without pre-processing or sanitization of potential injection vectors in HTML or text.

agent-browser