Real-time security firewall for LLM agents — blocks prompt injection, jailbreaks, data leakage and unsafe outputs across any model.