AVP – an agent can't leak a secret it never had

AVP is a security project that prevents API key leakage from coding agents by ensuring agents never have access to real secrets—only a proxy that validates requests before forwarding them. It targets developers running AI coding tools like Claude Code or Codex on their own machines, who need to grant agent access to APIs without exposing credentials. The project is interesting because it inverts the typical security model, making it impossible for agents to leak secrets they never possessed, even under prompt injection attacks.