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

Know the Harnesses

Part 1 argued the attacker changed. This part names the change, specifically: the agentic pentesting harnesses that are actually running against production systems right now, who is running them, and what they are good and bad at. A defender who cannot describe the opposing tooling in detail is planning against a rumor.

Start with the one that forced the issue. ARTEX is an open source autonomous penetration testing system: a planner agent decomposes a target into goals, worker agents claim tasks and execute them with real security tools, and a retest agent goes back over claimed findings to check whether they reproduce. It is a harness in the modern sense, a loop of models, tools, and verification wrapped around an objective, left running.

On November 13, 2026, the FBI issued a FLASH alert on it. The public record around that alert is unusually rich, because several independent parties were measuring the same phenomenon at the same time:

ARTEX is the biggest name, not the only one. The working set a defender should know includes Strix, PentestGPT-Agent, CyberStrike, Decepticon, LuaN1aoAgent, and Cairn, a dispatcher that orchestrates general coding agents like Codex and Claude Code as its attack workers. Cloudflare's top five by observed traffic: ARTEX, LuaN1aoAgent, Decepticon, Cairn, PentestGPT-Agent. The names will rotate. The architecture will not: planner, workers, tool access, a retest or verification step, and a console a human supervises loosely or not at all.

Two details from the ARTEX story deserve a defender's full attention, because both generalize.

The first is the trojanized copy. The most popular Docker image of ARTEX, pulled more than 8,800 times, carried a backdoor: an unauthenticated remote code execution flaw hidden behind a magic prefix in a login parameter. Anyone who ran the convenient image handed their attack infrastructure to whoever planted it. The registry deleted the image in early November; a byte-identical copy reappeared under a different account, and hundreds of forks had already inherited the poisoned reference. The lesson is not about one image. Offensive tooling is now valuable enough to be a supply-chain target itself, and the people downloading attack harnesses fastest are the least likely to build from source and check. If your own team runs these tools, and it should, build from pinned source, verify it, and treat the harness supply chain with the same suspicion you apply to your own.

The second is what the harness removed. In the public code, the rules-of-engagement machinery, the part that constrains what the system is allowed to touch, is gone. The destructive-action guardrails that remain are database rules seeded as ordinary configuration, which an operator can edit or delete. Read that as an engineering statement about where the speed comes from: part of the offensive advantage is subtraction. No scope review, no authorization gate, no change window. A defender cannot subtract their way to parity and should not try. But a defender who does not understand that the opposing system's speed is partly the absence of process will keep prescribing more process as the cure.

So what does a defender do with all this? Run the harnesses. Not read about them, run them, against your own systems, in a controlled range, with your own targets and your own telemetry watching. That exercise teaches things no report carries: which paths the planner tries first against your stack, which of your controls it routes around without noticing, where it stalls, what its traffic looks like on your sensors, how long a full run takes against an estate your size. Every one of those answers changes a defensive decision. The fingerprints alone are worth it: these systems have default ports, default user agents, characteristic request rhythms, and a defender who has watched a hundred runs in a range recognizes the first hostile run in production.

There is a standing objection to this, that defenders should not touch offensive tooling. It was wrong when the tooling was Metasploit and it is wrong now, at higher stakes. The harnesses are public. The attackers' copies are already running. The only question is whether the first time your organization meets one is in rehearsal or in an incident review.

Next in this series: the finding side of the loop. Scanning got so good that it broke the program downstream of it.