ARTEX and Friends: What Agentic Pentest Frameworks in Real Intrusions Mean for Defenders
The AI Pentest Agents series was written from the point of view of someone evaluating these tools: what do they do, where do they break, does the human pentester still have a job. Since 30 October that framing has a second half. Google’s Threat Intelligence Group and Mandiant published findings that open-source agentic pentest frameworks are being used by real threat actors in real intrusions, and ARTEX is the one most of the reporting is about.
This post is a deep dive on ARTEX and the other tools in the same space, and on what their misuse is doing to the defender’s job. It is deliberately not a lab write-up. I haven’t run ARTEX, the project is no longer open source, and I’m not going to publish a how-to for it. Everything below comes from vendor and researcher reporting, and I’ve tried to say whose claim each number is.
What changed on 30 October
Google and Mandiant reported that actors have been operating ARTEX, CAIRN and PentAGI in Mandiant incident response engagements since late August 2026, and that Hexstrike has also been abused by multiple actors. They compared the situation to Cobalt Strike and Metasploit: tools built for security professionals that attackers borrowed. The difference they stress is where the work gets done. In the cases they describe, agents carried out significant parts of the attack themselves, including lateral movement, database exploitation and exfiltration preparation.
The most useful sentence in their framing is one I’d have written myself: the techniques haven’t changed. Attackers are still using stolen credentials and unpatched systems. What has changed is how quickly and how widely an automated agent can work through them.
What ARTEX is
ARTEX is a multi-agent penetration testing framework written in Go. An operator gives it a goal, planning agents map the target and choose next steps, worker agents run scanners and exploit things, and the results feed back to the planners for the next wave. Early write-ups described a Next.js web UI and a dual-graph design, with a per-task exploration graph kept separate from a shared graph of discovered assets. Google describes the full attack lifecycle as split across six specialised sub-agents, though it doesn’t name them. It is model-agnostic: it talks to whichever LLM API the operator plugs in, frontier or open-source.
The history matters more than the architecture:
- A developer published it on GitHub in July 2026, pitched at learning and authorised testing.
- After it was used against South Korean financial firms, the developer closed the source on 8 October and stopped updates, citing misuse. Sources differ on whether the repository was closed or deleted outright.
- It didn’t go away. Forks survive, including a regional variant and a commercial version called ARTEX PRO.
Check Point’s later research found 2,557 systems that host or have hosted ARTEX control panels, 1,608 of them live and publicly reachable when they looked. That number is the clearest evidence that closing the source after the fact bought very little.
The tool landscape
ARTEX isn’t alone, and some of its neighbours already have posts on this site. This is what I can say about each, limited to what the reporting supports:
| Tool | What’s known | Where it’s been reported in the wild |
|---|---|---|
| ARTEX | Go, multi-agent planner/worker, model-agnostic, closed since 8 Oct, forks and a paid PRO edition | South Korean banks, a 200+ organisation campaign, Microsoft 365 tenants via third-party app identities, Check Point’s exposed operator consoles |
| PentAGI | Open-source multi-agent platform, covered in Part 4 of the series | Mandiant incident response since late August |
| CAIRN | Named by Google as an open-source autonomous pentest framework; Google gives no architecture detail, so I can’t either | Mandiant incident response since late August |
| Hexstrike | Open-source agent framework. In an earlier GTIG report it was paired with Graphiti, a temporal knowledge graph that kept a persistent model of the target, and switched between tools such as subfinder and httpx on its own reasoning | A Japanese technology firm (earlier GTIG report); multiple actors in Mandiant cases |
| Strix | Multi-agent pentest framework, used alongside Hexstrike in the same earlier GTIG report to automate finding and validating vulnerabilities | Same report as Hexstrike |
Two patterns stand out. First, every one of these is a legitimate pentest tool by design, so there’s no separate “criminal” product category to block. Second, the interesting component is increasingly the memory and orchestration layer (graphs, shared state, planner/worker splits), not the model itself. That is the same engagement-state idea the series found inside PentestCode and PentAGI, now being run by someone else.
Four campaigns, one pattern
The reporting breaks down into several clusters. Vendors name actors differently, and I haven’t verified which of these overlap, so treat them as separate until someone shows otherwise.
- A China-nexus cluster CrowdStrike calls Vapor Panda. Reported as a single ARTEX controller hitting 200+ organisations in at least 14 countries over 11 days, with post-exploitation confirmed at nine. These figures reach me via SOCRadar’s and The Hacker News’s summaries.
- The Storm-0940 / UNC7228 activity. Wiz tracks an ARTEX-related campaign that matches what GTIG calls UNC7228, which overlaps Microsoft’s China-linked Storm-0940. Google says it began at least as early as August. The reported method was abusing third-party application identities to reach Microsoft 365 tenants without interactive sign-ins; the tenant count in circulation is over 2,250. Google quotes about four hours on average from a first stolen account to deep privileges, and one intrusion that finished in fifteen minutes.
- A financially motivated actor against South Korean finance. CrowdStrike’s report covers late September to early October. The primary LLM backend was DeepSeek, reached probably through a reseller, with GLM and Grok also seen. Exposed directories on the actor’s infrastructure included Claude Code configuration files and session histories. The actor also used AI assistants for operational chores like asking where stolen Korean data is usually sold.
- A criminal ring, per Seoul police. Reported as seven people targeting 90+ organisations, breaching about 60 and taking roughly 8.2 million records, with the leader a former penetration tester. This one is from SOCRadar’s summary of the police disclosure.
Check Point added the most interesting dataset. Ten operator consoles were misconfigured to expose the full ARTEX database, which showed exactly how the operators prompted their agents, which models they used and what the agents did. Most attacks ran almost fully autonomously once a goal was set, with the operator mostly checking progress. In one case an operator pointed the agents at grey-market iOS app-signing services. Over about two weeks the agents collected more than 300 signing certificates with their passwords, and only three turned out to be usable. Reported costs for DeepSeek tokens across campaigns ran from roughly $145 to $1,500.
What the agents look like from the defender’s side
This is the part that connects back to Part 6’s fingerprinting question. That post predicted that default, unattended agents would be noisier than a human tester. Google’s ARTEX analysis supports it, with specifics:
- Burst timing. The gaps between phases look human, but execution bursts finish in roughly 2 to 25 seconds. In one intrusion, 27 security events landed inside a 23-second burst while the agent surveyed AWS permissions across security groups.
- Trial and error leaves evidence. One agent generated 240 web log events against an exposed Linux web server. Sub-agents overwrote each other’s files and commands, producing piles of error entries.
- A recurring artefact. Filenames following an
artex-probe-pattern with five-digit incrementing numbers, which also turned up in injected account names, passwords and email subjects. - Exfiltration through a webhook service. One agent rotated through 491 webhook.site URLs to get command output back.
- Scale. More than 110,000 Active Directory objects enumerated in one case.
The tradecraft underneath is the stuff this site has been writing about for months: Kerberoasting, DCSync, password spraying, delegation abuse. The Impacket detection post covers the host and network side of much of that. None of it needs AI-specific detection logic. It needs the existing detection to be actually deployed.
What I’d do about it
Mandiant’s guidance is five layers: protect the control plane, minimise accounts and long-lived secrets, harden configuration and reduce attack paths, put guardrails on high-impact application APIs, and build behavioural baselines for burst rates and unusual egress. Google itself describes this as established best practice. As a network engineer, the parts I’d act on first:
- Take management planes off the internet and separate them. Admin APIs, consoles and CI/CD interfaces on trusted networks only. I covered one way to do this on FortiGate in the management plane separation post.
- Filter egress to webhook and paste-style services you have no business use for. An agent that can’t get output out is a much less useful agent.
- Alert on velocity, not just content. Many events from one identity in seconds, or dozens of AD queries in a burst, is the signal that a person typing doesn’t produce.
- Hunt for the artefact string. If you run Wazuh, a rule on the
artex-probe-pattern in web logs is an obvious first idea. I haven’t built or tested one, so treat that as a suggestion, and see how a log line becomes an alert for the mechanics. - Export logs before you delete anything during a response. Sign-in, audit and SIEM logs are the evidence trail.
What this means for the industry
Dual-use isn’t new, tempo is. Cobalt Strike and Metasploit were always borrowed. The change is that the operator doesn’t need the skill to drive them. Check Point’s reading is that these systems let attackers work at a speed that wasn’t previously possible and without technical knowledge. The cost figures above support that: a campaign can now be run on an API key and a token budget of a few hundred to a couple of thousand dollars.
Patch windows shrink. If one intrusion can go from a stolen account to deep privileges in 15 minutes, a patch cycle measured in weeks is the wrong shape. This is a bigger point than any single tool.
The attackers have the judgment gap too. The Judgment Gap post was about defenders drowning in unvalidated AI findings. Check Point’s certificate case is the mirror image: hundreds of stolen certificates, three of them working, and an operator getting confident progress messages from agents that were mostly wrong. Agents overstating success is a property of the technology, and it applies to whoever is running it.
Closing the source didn’t contain it. Forks and a paid edition circulate, and more than 1,600 panels were still reachable when Check Point looked. Ten of those were misconfigured badly enough to expose their full databases, which handed researchers a rare view of how operators actually use these tools, and gives defenders something to learn from.
It doesn’t make human pentesters obsolete. If anything, the picture is the same as the one in the series: the agents are fast, tireless and noisy, and a human who can validate, scope and interpret is still where the value sits. The difference is that the same speed is now available to people with no authorisation at all.
What I haven’t tested
I haven’t run ARTEX, CAIRN, Hexstrike or Strix for this post, and with ARTEX closed-source I haven’t been able to read what it actually does beyond what researchers have published. The architecture description is second-hand. The campaign figures are the reporting vendors’ claims, several of them reaching me through summaries rather than the original reports, and the same actor can carry different names depending on who is writing. This is also a fast-moving story, so what’s above is a snapshot as of 10 November 2026. If you spot something I’ve misattributed, tell me and I’ll correct it.
Sources
- Google Cloud blog, “ARTEX and Friends: Enter Agentic Offensive Frameworks” (GTIG and Mandiant, 30 October 2026), Korean-language edition
- CrowdStrike, unknown threat actor uses ARTEX to target South Korean finance
- SOCRadar, ARTEX AI Pentest Tool: What Security Teams Need to Know
- Check Point Research, “Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns” (5 November 2026), read via GridinSoft’s summary and Let’s Data Science
- Let’s Data Science on the Google and Mandiant report
- Earlier GTIG reporting on Hexstrike and Strix: The Next Web