If you discover a security vulnerability in the CUP Python SDK, please report it responsibly.

Do not open a public issue.

Instead, email cup@computeruseprotocol.com with:

• A description of the vulnerability
• Steps to reproduce the issue
• The potential impact
• Any suggested fixes (optional)

We will acknowledge receipt within 48 hours and aim to provide an initial assessment within 7 days.

CUP interacts with OS accessibility APIs and can execute UI actions on behalf of AI agents. Developers integrating CUP should be aware of:

• Action execution scope — CUP can click, type, and interact with any accessible UI element. Always validate and constrain which actions an agent is allowed to perform.
• Element references — Element IDs are ephemeral and scoped to a single tree capture. They cannot be reused across sessions.
• MCP server — The cup-mcp server runs locally over stdio. It does not expose a network interface by default.
• No credential storage — CUP does not store, transmit, or handle credentials. API keys for LLM providers belong in your application layer, not in CUP.

• Run CUP with the minimum OS permissions needed for your use case
• Implement action allowlists in your agent layer — don't give agents unrestricted access to all CUP actions
• Review tree captures before sending them to external LLM APIs, as accessibility trees may contain sensitive on-screen content
• Keep CUP updated to receive security fixes