Every line below is a real command run against the actual extracted detector and TOTP-verification code, in an isolated sandbox — not a mockup of what we expect to happen. This is the same methodology we'd use to evaluate an adaptive attacker or an autonomous agent probing the box: start from observed behavior, not from reading the source.
1,000 source IPs, 2 attempts each. Grouped by busiest single IP — fired.
New distributed-burst tier catches it: 1,000 failures across 500 sources inside the window.
600 sources probing one unexpected port. Counted globally — caught correctly.
480 connections/30s, just under the burst line, sustained. The long-window tier catches it 12 minutes in.
Blank, malformed, and wrong codes all fail closed before the real check even runs.
Both detectors match IPv4 patterns only — an IPv6-sourced flood currently goes counted.
Why this matters for autonomous/agentic attackers specifically: a scripted brute-forcer gives up after a fixed number of tries. An adaptive agent calibrates — it backs off when it gets cut, and creeps its rate back up. The long-window tier exists for exactly that behavior: a fixed burst limit alone is beatable by patience, so there's a second counter with a longer memory. Nothing here claims that's unbeatable — see the security page for the honest ceiling on that defense.