If you discover a security vulnerability, please report it responsibly:

1. GitHub: Open a private security advisory
2. Expected response: Within 72 hours

Please do not open a public GitHub issue for security vulnerabilities.

In scope:

• LLM prompt injection that causes unintended tool execution
• Path traversal in tool parameters (e.g., reading/writing files outside compositions/)
• Subprocess injection via malformed tool arguments
• Credential exposure in logs or error messages

Out of scope:

• AI model output quality or bias
• Denial of service via large generation requests (local tool, user's own hardware)
• Issues in upstream model repositories (ACE-Step, NotaGen, etc.)