Add comprehensive security acceptance testing framework that validates
Moltbot's resistance to prompt injection, data exfiltration, and trust
boundary violations.
Key components:
- LLM-as-judge pattern using Claude to evaluate attack resistance
- WebSocket gateway client for direct protocol testing
- CLI mocking utilities for injecting poisoned external data
- Docker Compose setup for containerized CI execution
- GitHub Actions workflow with daily scheduled runs
Test categories covered:
- Email/calendar prompt injection via external data
- Trust boundary violations and auth bypass attempts
- Data exfiltration prevention
- Tool output poisoning
