ARTEX: The FBI's Warning About AI Agents That Run the Whole Intrusion
The FBI's November 13, 2026 FLASH warns that ARTEX, an open-source agentic AI penetration testing framework, is being used to go from first scan to privileged access in under three hours. How it works, the South Korea and ConnectWise campaigns, the intrusion chain, what to look for in your logs, the FBI's mitigations and three things to do today.
Watch the video version on YouTubeIn this deep dive
- What ARTEX is: a Go orchestrator that hands an intrusion to AI agents, and why pulling the GitHub repo on October 8 didn't stop it
- How it works: SI agents in parallel, reasoning rented from Claude, GPT, DeepSeek, GLM, Grok or Kimi, often through API resellers
- The FBI's five risks, from multi-target operations to a compressed detection window
- Two campaigns: about 67,000 people's data from South Korean financial institutions, and 2,250+ tenants reached through ConnectWise Cloud Backup's OAuth app
- The intrusion chain, the lateral movement EDR blocked, and why the initial access flaws were all testable
- Indicators and three places to detect it: identity, egress to LLM APIs, and the cloud control plane
- The FBI's mitigations, what it means for Security+, CySA+ and SecAI+, and three actions for today
Source: FBI FLASH-20261113-001, "Threat Overview of ARTEX Autonomous Penetration Testing Framework," November 13, 2026 (TLP:CLEAR). The full write-up, with the ATT&CK mapping, the IOCs and both hunting queries, is on the Skillthropic blog.
Links
Transcript
On November 13th, the FBI published a FLASH warning. Its official name is FLASH 2026-1113-001. There's a mouthful. It's marked TLP:CLEAR, which means it's meant to be shared, and it's about a tool called ARTEX. ARTEX is a penetration testing framework that hands an intrusion to a team of AI agents. You give it a goal in plain English, connect it to a large language model, and it plans and runs the reconnaissance, the exploitation, the privilege escalation and the data collection on its own. The FBI says threat actors have used it since at least August, and some of those intrusions went from first targeting to privileged access in under three hours. We're going to walk through what ARTEX is, how attackers are using it, the intrusion chain, what to look for in your logs, what the FBI recommends, and three things you can do today.
ARTEX is an autonomous, multi-agent penetration testing platform written in Go. A developer using the handle Autumn-27 released it on GitHub in July of 2026, for authorized testing. It picked up an estimated fifteen hundred to two thousand stars in about ten weeks. After reports of abuse, the developer pulled the repository on October 8th. Did that stop it? No. The FBI notes that ARTEX keeps spreading through community forks, localized variants, archived copies, and a commercial edition called ARTEX PRO. Researchers counted more than sixteen hundred publicly exposed ARTEX consoles in early November. Once a tool like this is out, it's out.
The design is what makes it dangerous. ARTEX has no model of its own. It's an orchestrator. It takes a high-level objective, breaks it into tasks, and assigns those tasks to specialized agents. The FLASH calls them SI agents. Each one turns its task into real commands, and they run in parallel: one doing recon, one exploiting, one working the cloud, and of course one working Active Directory. The reasoning comes from whichever model the operator plugs in. The FBI lists frontier models like Anthropic's Claude and OpenAI's GPT, and open or Chinese-origin models like DeepSeek, GLM, Grok and Kimi. In attacker-run instances, the Chinese models dominate, led by the DeepSeek Flash family. Why? Well, they're cheap, they're capable, and the FBI saw no observable guardrails. Operators often reach them through API resellers, so the traffic doesn't always go straight to the model provider. You can think of ARTEX as a general contractor. It doesn't swing the hammer itself. It reads the plans, hires the crews, and keeps every crew working at the same time.
The FLASH describes ARTEX less as a new exploit and more as a force multiplier. Nothing in its toolbox is new. SQL injection, SSRF, Kerberoasting, web shells. These are decades old. What changes is who can chain them together, and how fast. The FBI calls out five risks. First off, accelerated multi-target operations. The agents automate the slow parts of an intrusion, enumeration and exploitation, so a small group can run several campaigns at once. Second, overlooked attack surfaces. ARTEX is good at finding the systems nobody watches: partner and supplier portals, mobile backends, employee apps, legacy services, and the app a business unit stood up on its own. Third, supply chain compromise. The agents map trust relationships between organizations and go through the weakest one. In one case, they modified iOS app distribution infrastructure using exposed cloud credentials. Fourth, a lower barrier to entry. Operators with minimal skill can run intrusions that used to take an experienced team. And then lastly, a compressed detection window. ARTEX is noisy. It does trip alerts. The problem is it often reaches its objective before a defender has finished triaging them.
Two campaigns show the scale. In the first, a financially motivated actor used ARTEX against South Korean financial institutions and exfiltrated personal information on about sixty-seven thousand people. The second is the one I'd want every Microsoft 365 admin to understand. It was a supply chain intrusion against ConnectWise Cloud Backup customers, and it opened access to more than two thousand two hundred fifty downstream tenants. How does one vendor turn into two thousand tenants? Well, the vendor had a multi-tenant OAuth application, and every customer had already granted that application admin consent. The attacker compromised the vendor's credentials, or added a new credential to the vendor's application, and that application already had a key to every customer's front door. Nobody phished the customers. The attacker inherited the access they'd already given their vendor. That's why so many of the FBI's mitigations are about Microsoft Entra and app consent.
The FLASH maps what ARTEX agents did across the whole intrusion lifecycle, and if you've studied for Security+ or CySA+, a lot of it will sound familiar. Reconnaissance: the agents mapped the permissions of stolen Azure credentials with test writes timed in fractions of a second, tested AWS keys with DryRun calls, and fuzzed more than five hundred web paths. Initial access: unsigned SAML assertions, XXE, SSRF, server-side template injection, SQL injection, insecure direct object references, APIs with no authentication, and admin passwords hardcoded in single-page-app JavaScript. Foothold: web shells, commands through Kudu admin APIs, open-source remote access tools, and new credentials on existing OAuth apps. Privilege escalation: Kerberoasting, DCSync of the krbtgt key, and admin tokens forged with keys nobody had rotated. Internal recon: more than a hundred and ten thousand Active Directory objects enumerated. Then collection, with data staged on new hundred-gigabyte volumes and sent out to rotating webhook URLs.
Two details matter more than the rest. First, when the agents tried to move laterally, with pass-the-ticket, pass-the-hash and WinRM, endpoint protection blocked them. Their response was to inventory which Windows hosts didn't have an EDR sensor. Coverage gaps are now something attackers enumerate automatically. Second, almost every initial access technique on that list is a web application or identity flaw that a code review, a secrets scanner or a SAML configuration check would catch. ARTEX isn't beating good controls. It's finding the places where controls were never applied. Five locks on the front door don't help when the side window is wide open, and ARTEX checks every side window you have.
The FBI does publish IP addresses and a domain, an LLM API reseller called xcai dot pro, and they're in the blog post. Keep in mind they're historical. A lot of that infrastructure sits on cloud and VPN hosts that may have been reassigned since, so corroborate any hit with your own telemetry before you block it or declare an incident. The behavioral indicators will outlive every IP on that list. Outbound calls to LLM APIs from systems that have no business making them. High-frequency automated requests with sequential enumeration against login and search pages. Structured personal data leaving over HTTPS to destinations nobody approved. Dot-aspx or dot-jsp files in places they shouldn't be. And the fingerprints of an agent swarm: compressed bursts of activity, lots of failed attempts with small changes between them, and duplicate or conflicting requests where sub-agents overwrite each other's work.
Agentic tooling is fast, but it isn't quiet, and that's our opening. Three places to look first. One, identity. Look for new credentials on applications and service principals. Adding a secret or a certificate to an app that already has broad consent was the quietest persistence in the ConnectWise campaign. In the blog post there's a Microsoft Sentinel query that pulls every service principal credential addition, every certificate and secret change, and every consent grant since October 1st. Review every row. Then look at service principal sign-ins, successful and failed separately, from anywhere outside the vendor's expected infrastructure. Two, egress. A web server, a database host or a build agent calling a model API is worth a question, even when it turns out to be legitimate. The second query in the post, for Defender XDR, counts connections to those LLM endpoints by device and by process. The rule of thumb is to build an allow list of the hosts and processes that are supposed to use AI services, and alert on everything else. Three, the cloud control plane. ARTEX tests stolen cloud credentials in a pattern people don't produce: test writes and deletes a fraction of a second apart, deployment dry-runs across many resource types, AWS calls with the DryRun flag, unexpected role assumptions. Alert on the burst, not the individual call.
The FLASH groups its recommendations into four areas. Reduce the external attack surface. ARTEX wins on the systems you forgot, so the first job is knowing what you expose. Inventory every internet-facing application, including the partner portals and the apps a business unit built on its own. Put cloud administration, Kudu and SCM publishing endpoints, hypervisors and backup platforms behind trusted networks. And require an architecture review before anything gets a public IP. Harden identity and access. Least privilege on multi-tenant app permissions and consent grants. If you ever installed an affected ConnectWise application, review every tenant where it is or was authorized, even if you stopped using it, and search by application ID, not display name. Display names aren't unique. Application IDs are. Strengthen application protections. Rate limits and transaction thresholds on high-impact APIs. Server-side authorization on every record. Stop returning stack traces and SQL errors to people who aren't logged in. Every error message is a hint to an agent that learns from feedback. And scan your repositories, images, pipelines and infrastructure code for embedded secrets, continuously. And then lastly, preserve evidence and respond. Keep your Entra, Purview and SIEM logs, but don't hold up urgent containment while the exports run. For ConnectWise-related activity, review back to at least October 1st. And don't delete the enterprise application as your first step. Remove the unauthorized credentials and grants after the evidence is preserved.
For Security+, on SY0-701 or SY0-801, ARTEX is a case study in attack surface reduction, application attacks and supply chain risk, and the SY0-801 draft adds an objective on AI threats. For CySA+, the FLASH is a ready-made exercise: take each behavior, map it to an ATT&CK tactic, then write the detection. And for SecAI+, this is AI-enabled attack automation in the wild, and the flip side of it: an agent's permissions decide its blast radius.
Three things to do today. First off, close the admin side doors. Find every Kudu or SCM publishing endpoint, backup console, hypervisor and management interface reachable from the internet, and move it behind a trusted network. Second, audit your app credentials and consent grants back to October 1st, and turn that query into a standing alert. And then lastly, put a speed limit on your partner, customer and employee portals, and alert when a server that isn't on your AI allow list calls an LLM API.
ARTEX did nothing a skilled human team couldn't do. What changed is the economics. An attacker with little skill and a cheap API key can probe every forgotten portal you own, in parallel, and keep going while your team sleeps. The answer isn't an AI arms race. It's closing the basic gaps on the systems nobody was watching. The full write-up is on skillthropic.com, with the ATT&CK mapping, the indicators, both hunting queries and the FBI's mitigations. While you're there, have a look at the study guides for Security+, CySA+ and SecAI+.
Enjoyed this? Get the AI security news that matters.
Join The AI Security Brief for the top AI security news, plus what's important to the C-suite. Free, straight to your inbox.
No spam. Unsubscribe anytime.
Follow Skillthropic
Keep learning between study sessions
The same exam-focused teaching, in smaller pieces, wherever you scroll.