Substack mockup · not published
Brandon Gilchrist · Defense at Machine Speed

Baby Steps

Airlines are getting hit right now. Air traffic control is getting hit. The NSA is getting hit. Banks and credit card companies are getting hit.

This is not a bad quarter. It is a structural change in what an attacker is.

The attacker is no longer a person at a keyboard working a target list at human speed. The attacker is a harness: an autonomous loop that scans, probes, chains identities, and retries around the clock without getting tired, without getting bored, and without waiting for your change window. In November the FBI issued a FLASH alert on one of these harnesses, ARTEX, an open source agentic pentesting system that has forked into dozens of variants. Cloudflare, measuring its own network, reports requests from AI penetration testing orchestrators up 30x since October 1, with ARTEX and its forks accounting for more than 60% of the known orchestrators. Security researchers counting exposed consoles found more than 2,300 offensive AI control panels sitting on the public internet. State-aligned actors and criminal groups are already running these systems in live operations, in one reported case against more than 200 organizations in 14 countries in 11 days.

That is the offense now. Here is the defense.

Most security programs still run on a calendar built for the old attacker. An annual penetration test. A quarterly scan. A remediation queue measured in sprints and quarters. A DAST tool configured to simulate a human tester clicking through an application at human pace. Every one of those rhythms assumes the person on the other side sleeps, specializes narrowly, and attacks one target at a time.

We have taken baby steps beyond preparing for a traditional attacker to attack us in a traditional way at a traditional speed. That sentence is the whole problem, and this series is about what replaces it.

Three things changed, and defense has to change with each of them.

First, finding vulnerabilities got cheap. Frontier models now scan at a scale no human team matches. Anthropic's Project Glasswing put its Mythos model against the codebases of roughly 50 partners and surfaced more than 10,000 high and critical vulnerabilities. The same report carries the number that actually matters: more than 99% of the candidate findings were still unpatched. Scanning is close to solved. Remediation is the bottleneck, and a defense program that still measures itself on findings is measuring the wrong end of the pipeline.

Second, the attack tooling is public, and defenders are allowed to use it. The harnesses hitting production systems right now are open source or one step removed from it. A defender can download the same systems, run them against their own infrastructure in a controlled range, and learn exactly what is coming: which paths the harness tries first, where it gives up, what it reports. Defenders need free and unfettered access to the same tools the bad guys use. That has always been true. At machine speed it is the difference between rehearsing against the real attack and rehearsing against a memory of one.

Third, defense has to become a loop, not an event. Scan everything with the best models available, now. Route the fixes through AI-assisted remediation with human review, because the volume no longer fits a ticket queue. Then scan again, and treat the second scan as the one that counts: a fix is real only when the exploit path fails on a re-check. Attack success rate, detection and containment time, fix survival under re-attack, cost per resolved incident. Those are the numbers of a defense program. Everything else is a report about one.

This series will work through the whole picture. The attack harnesses, one by one, with receipts: what they do, who runs them, where they are weak. The defense theories that hold up at machine speed and the ones that do not. Deception that prices defense in attacker tokens and time. And the operating model underneath: continuous scanning, verified remediation, and a range where your own systems face the real harnesses before the internet's copies do.

The attackers already made the transition. The only open question is whether defense makes it on purpose, or keeps taking baby steps while the incidents compound.