A companion to Part 2 of Defense at Machine Speed. The main piece argues defenders should know the agentic attack tooling in detail. This is the detail: each harness in the current working set, what it is, what the public record supports, and what a defender should take from it. Where the record is thin, the profile says so. Thin is a finding too: it means the range, not the literature, is where that harness gets understood.
The center of gravity. An open source autonomous penetration testing system, written in Go, organized as a set of cooperating agents: a goals decomposer breaks a target into objectives, a planner generates the next intent, workers claim one intent at a time and execute it with a real tool arsenal, and a retest agent revisits claimed findings and renders verdicts, reproduced, fixed, or inconclusive. One database holds both a per-task exploration graph and a global asset graph, so what one run learns about a target persists into the next.
The record: FBI FLASH alert, November 13, 2026. Cloudflare counts 12,000+ repositories in the family and ARTEX plus forks at over 60% of known orchestrator traffic. Spectrum Security catalogues 47 variants and 2,300+ exposed consoles. ReversingLabs resolves the ecosystem into 11 fork families over one shared core. Microsoft, Wiz, CrowdStrike, and Unit 42 reporting ties it to a China-aligned actor active since September 2026, a crew that hit 200+ organizations in 14 countries in 11 days, and a derivative, AntiFox, observed stealing cloud credentials in live intrusions. The most-pulled Docker image carried a backdoor; hundreds of forks inherited the poisoned reference.
Defender's notes: the public code's scope-enforcement machinery is removed, and its remaining guardrails are editable configuration, so assume any hostile copy runs unconstrained. Its agents run with broad tool permissions; containment is entirely a property of the deployment, which is why your own runs belong in an isolated range. Build from pinned source, never from a public image.
A dispatcher, not a monolith: Cairn orchestrates general-purpose coding agents, Codex, Claude Code, and Pi among them, as its attack workers. That design is the point. It does not need its own exploitation models; it rents the frontier's coding ability and supplies the loop, the tasking, and the persistence around it.
The record: top five in Cloudflare's orchestrator traffic table, November 2026. Beyond that, the public record is thinner than ARTEX's: fewer independent teardowns, less incident attribution.
Defender's notes: Cairn-class systems mean the capability ceiling of the attack is the capability ceiling of the best coding agent, which moves every model release. Evaluating one in the range is straightforward because its workers are the same agents your own engineers may run. Watch what it does to your assumptions about agent traffic: its workers look like developers, because they are developer tools.
The agentized descendant of PentestGPT, the assistant that guided human operators through penetration tests. The agent version closes the loop: the system that used to advise the operator now plays the operator.
The record: top five in Cloudflare's traffic table. Public detail beyond that is limited, mostly the lineage from the assistant era.
Defender's notes: the lineage is the lesson. Every advisory tool in this space is one engineering step from autonomy, and the installed base of people who learned the methodology from the assistant version is the distribution channel for the agent version. Expect competence at standard methodology, executed tirelessly.
The least documented name in Cloudflare's top five, which is itself worth sitting with: a harness can reach the top of observed orchestrator traffic while leaving almost no public analytical footprint in English-language sources.
The record: second in Cloudflare's top-five table, November 2026. No FBI alert, no major vendor teardown in the public record as of this writing.
Defender's notes: treat it as the standing reminder that the visible ecosystem, the one with alerts and blog posts, is a sample, not the population. The range program should be built to evaluate a harness from its behavior, quickly, because the next important one will also arrive undocumented.
Named in Cloudflare's top five and in the current working set of harnesses defenders are tracking. Public technical detail is thin: no widely circulated architecture teardown, no dedicated alert.
The record: top-five observed traffic, Cloudflare, November 2026. That is close to the whole of it.
Defender's notes: same conclusion as LuaN1aoAgent, and the pair of them together make the point better than either alone. Two of the five most active orchestrators on the world's largest network are near-blank pages in the public literature. Measurement is ahead of understanding, and the defender's own range work is how understanding catches up.
An open source agentic penetration tester that has circulated in the security community as a usable tool rather than a research demo. It appears in OrcaRouter's Shield testing set, the group of harnesses a deception vendor chose as representative opponents: ARTEX, Strix, PentestAgent, CyberStrike, and Cairn, all run against a frontier open model.
The record: named in vendor testing and community use; no dedicated alert or major incident attribution in the public record as of this writing.
Defender's notes: Strix is a good first range subject precisely because it is a community tool: obtainable, runnable, and representative of the class a mid-skill operator can deploy today. The gap between Strix-class tooling and ARTEX-class operations is mostly scale, supervision, and target access, not kind.
The second name in this set known mainly through its appearances in other people's measurements: Cloudflare-era tracking lists and OrcaRouter's Shield test set.
The record: named as a representative offensive harness in vendor testing, November 2026. No public architecture teardown, no incident attribution on record.
Defender's notes: file with LuaN1aoAgent and Decepticon. The pattern across this companion is the finding: of seven harnesses a defender should know, one has a full public record, two have partial records, and four are mostly names in traffic tables. That ratio is the state of the art in defender knowledge, and closing it is unglamorous, necessary work: run them, watch them, write down what they do.
Read as a group, the seven split into three kinds: full autonomous systems with their own agent architecture (ARTEX), dispatchers that weaponize general coding agents (Cairn), and a long tail of tools known by reputation and traffic rather than documentation. All three kinds converge on the same loop: plan, execute with real tools, verify, persist what was learned. A defense built for that loop, the subject of the rest of this series, does not need to re-tool for each new name. It needs the range time to keep confirming that the loop it faces is the loop it rehearsed against.