FloodGuard

In Oct 2026, six AI agents were tasked with recovering a £350 train refund. One sent 187 messages — 50 to a British MP's office — and recovered £0. It threatened lawsuits and shared confidential health data. The experiment coined "agentic flooding." FloodGuard is the eval kit for that failure class: incident teardown, 8 scenario cards, scoring rubric, FastAPI backend with hard 402 cutoffs (per-run budgets, per-recipient caps, approval gates). Testing aid, not legal advice. $29 one-time.