On a November afternoon I built a fake attacker and knocked on my own server thirty times. Thirty fast requests at wp-admin, phpMyAdmin, .env — the standard scanner routine. Ten seconds in, the server stopped answering that address completely. No ticket, no human at 3 a.m.: a firewall DROP rule with no undo. The attack was over before I finished writing the test plan.

The tool is CrowdSec, wired into my five public sites this past Sunday. This is a post about timing — what the agentic era stole from every defender. The fix is not to become faster. The fix is to stop requiring you.

Machine-speed offense

My last post, in brief: when PHP2Shell — two 9.2-rated PHP CVEs — landed in October, exploitation was attempted four days before the patch existed. An advisory diff is homework for every exploit agent on earth, done in minutes. The attacker’s bottleneck was never cleverness; it was effort, and effort has been automated away.

Our side still runs the human loop — log, alert, triage, decide, block — and that loop is hours even when it works. At agentic volume the alert queue becomes the attack: machine-speed noise is denial of service on attention. Reacting is not losing the race; it is agreeing to a race you cannot win. The posture that survives is pre-action. My last post argued removal — static sites, bugs with nowhere to live. This is the other flavor, for surfaces you cannot remove: a response that is pre-authorized, machine-executed, and done before you are involved.

What CrowdSec actually is

An open-source collaborative intrusion prevention system. Three parts. Parsers tail your logs — Caddy, nginx, SSH, hundreds more — and structure them. A scenario engine decides what is suspicious: brute force, 404-spam, admin probing, named CVE probes. The firewall bouncer is the blunt instrument I care about: a ban becomes a DROP in your iptables before a packet reaches an application.

Squint and it is fail2ban with a brain. The differentiator is “collaborative.” An opting-in install feeds its verified decisions to a central hub and pulls back everyone else’s. An agent that torched a thousand homes abroad arrives pre-banned at yours. Not threat intel sold after the fact — herd immunity as architecture. The day I switched mine on, the ban list filled with attackers other people had already burned. I didn’t ask for that. It just arrived in the ipset.

The build: one container and one process

My public surface is five sites behind Caddy on a Unraid NAS. The first layer is static: geo-fences — Canadian allow-lists evaluated at the edge, because if your users are one country, the rest of the internet is not customers, it is volume. But a fence cannot react — a VPN inside the allowed countries knocks forever. CrowdSec makes knocking expensive.

One container tails the Caddy JSON logs I already had writing; the whitelist drops my LAN, VM, Tailscale and Cloudflare ranges — you never ban your own infrastructure. The bouncer is deliberately not a container, because the point is host iptables: it hooks INPUT and DOCKER-USER, so host services and container ports die on one ban. One site rides Cloudflare, so the logged peer is the CDN itself — ban the peer, ban your own site. Strict trusted-proxies plus CF-Connecting-IP restores the real visitor, and I proved the spoof dies at the CF edge. Automatic reaction is a loaded gun; the whitelist is the safety.

A swarm of cold blue eye-like orbs on black tendrils assaults a monumental gateway sealed by a golden barrier; the attackers shatter into golden embers at the threshold.
Five sites, one fence, one bouncer — probing agents shatter at the edge before anyone gets paged.

Proving it to myself

Trust is measured, not configured. A synthetic attacker: a network namespace wearing a TEST-NET address inside the Canada allow-list — the fence waves it through — thirty fast probes at admin paths. The probing scenarios fired; a four-hour ban was enforced inside ten seconds. The next request got nothing back: not 404, not 403, no answer. A control address, same probes, 200s all the way. Log, parse, decide, ban, propagate — end to end, without me.

Why proactive, why now

Agentic AI did not make attackers cleverer; it collapsed the cost of attack. Scanning four billion addresses forever is free, and an exploit runs identically against a bank and a NAS in somebody’s basement. Defense wins on two levers. Lever one: not be there — removal, my last post. Lever two: make attacking you unprofitable. Every probe that costs the attacker a ban at your edge makes scanning your address worth zero, and the agent moves on. That is not a moral victory; it is pricing. Attention is the one resource you cannot scale, and a human bottleneck at machine speed is itself a vulnerability. Automate the boring verdicts, reserve humans for what only they can judge.

The honest limits

A ban patches nothing. A CVE in your request path still ends in removal or a patch — CrowdSec buys hours, not immunity — and an attacker on fresh residential proxies looks like a visitor to any firewall, which is why the fence, the auth, and the static-first architecture carry their own weight. And automation bites: a bad scenario can ban a real user, which is why my bans expire. My mistakes evaporate by morning; the script kiddies’ don’t.

The posture

Ordered, for anything listening on the internet: remove before you harden — every interpreter deleted is a CVE you never chase, the argument is here. Deny narrowly — if your users are in one country, that country is the allow-list. Then automate the response at the edge — bans in ten seconds, expiring, no analyst. Join the herd: your logs are worthless alone and an immune system shared. And write the never-ban list before the ban-everything list.

The machine that served this page bans people. Between the headline and this paragraph, agents have already probed it — and some were answered by the firewall before I knew. Defense in the agentic age is not about being smarter than the attacker. You will not be. It is about not needing to be awake.

— Kris

Permalink