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

Baby Steps (evidence-led opening)

On November 13, 2026, the FBI issued a FLASH alert about a piece of open source software. Not a vulnerability. Not a crew. A harness: ARTEX, an autonomous penetration testing system that plans its own attacks, executes them with real security tools, and retests its own findings before reporting them.

The alert was not the only measurement that week. Cloudflare, counting traffic across its network, reported requests from AI penetration testing orchestrators up 30x since October 1, with ARTEX and its forks responsible for more than 60% of the known total, in a family tree of more than 12,000 repositories. Researchers cataloguing the ecosystem found more than 2,300 offensive AI consoles exposed on the public internet. And the incident reports kept landing in the sectors that cannot afford them: airlines, air traffic control, the NSA, banks, credit card companies, all getting hit in the same season.

Put those two facts together and the shape of the moment resolves. The offense industrialized. It became a loop that runs around the clock, plans, probes, chains identities, and retries without fatigue or working hours. The defense, in most organizations, did not industrialize with it. It still runs on a calendar: an annual penetration test, a quarterly scan, a remediation queue measured in sprints, a DAST tool pretending to be a human tester clicking at human pace.

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 thesis of this series, and everything in it is an attempt to close the gap it names.

Three changes drive the argument. First, finding vulnerabilities got cheap: Anthropic's Project Glasswing surfaced more than 10,000 high and critical vulnerabilities across roughly 50 partners with its Mythos model, and reported more than 99% of the candidate findings still unpatched. Scanning is close to solved; remediation is the bottleneck. Second, the attack tooling is public, and defenders can and should run the same harnesses against their own systems in a controlled range, learning the real attack instead of a memory of one. Third, defense has to become a loop: scan everything, fix with assistance, scan again, and count a fix only when the exploit path fails on a re-check, measured on attack success rate, containment time, fix survival, and cost per resolved incident.

The parts that follow work through each piece with receipts: the harnesses one by one, the economics of scanning and fixing, the verification discipline, the deception layer that taxes the attacker's loop, and the operating model that assembles all of it.

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.