Continuous adversarial-testing (proof) engine for AI-agent defenses — run curated attack packs against the deployed stack, verify enforcement with the kill-chain, and publish an honest, reproducible defense benchmark.