Sunday, November 15, 2026

The FBI Just Warned About a Pentest Tool. Here Is Why ARTEX Is Different.

On November 13, the FBI released a FLASH alert, FLASH-20261113-001, about a penetration testing tool. That is unusual. The Bureau does not normally warn critical infrastructure operators about pentest frameworks, because pentest frameworks are supposed to be run by the defenders who hire pentesters. This one, called ARTEX, stopped staying on that side of the line months ago.

ARTEX is an open-source, multi-agent framework written in Go with a web front end. A developer using the handle Autumn-27 published it to GitHub in July and took the repository down on October 8 after reports of abuse. Taking it down accomplished very little. By early November, researchers counted more than 1,600 ARTEX consoles exposed on the public internet, and the tool lives on through mirrors, community forks, localized builds, archived copies, and a paid version called ARTEX PRO. Code does not come back once it is out, and this code was useful to exactly the people you would not want to have it.

What makes an agentic framework different

Plenty of tools automate parts of an attack. A scanner enumerates open ports. A password sprayer tries one password against many accounts. What those tools do not do is decide what to try next. A human operator reads the scan, forms a theory, picks the exploit, interprets the result, and chooses the follow-up. ARTEX moves that judgment loop into software.

It runs two kinds of agents. Planning agents map the target and choose the next step. Worker agents run the offensive tooling and feed the results back up. The planning agents read those results, revise the plan, and launch the next round, keeping a graph of every asset and weakness they have found along the way. The framework can pursue several breach paths at the same time and does not get tired, distracted, or nervous. Microsoft describes activity running from reconnaissance through data exfiltration at what it calls machine speed, with some intrusions reaching privileged access in under three hours.

The intelligence comes from whatever model you point it at. ARTEX works with any large language model, and operators in reported campaigns have wired it to DeepSeek, GLM, and Grok. Behavioral detection guidance from Zscaler, cited in the FBI alert, lists unexpected outbound calls to LLM endpoints such as api.anthropic.com and api.deepseek.com as a signal that something on your network is thinking out loud to a model it should not be talking to.

From tool to campaign

The reason the FBI, Microsoft, and Sophos are all writing about the same tool in the same three weeks is that it showed up inside a real intrusion with real consequences. Microsoft attributes the mid-October ConnectWise Cloud Backup compromise to a China-state aligned actor it tracks as Storm-0940, a group active since 2021 and previously known for grinding password spraying and OAuth abuse. Storm-0940 used ARTEX to run that old playbook far faster than it used to.

The numbers give a sense of the tempo. Between October 10 and 14, a single piece of infrastructure sprayed passwords at 327 accounts across eight organizations, generating more than 151,000 failed logons and peaking at 2,900 failures per minute on October 11. After the ConnectWise compromise, the actor used stolen credentials from ConnectWise cloud applications to authenticate into customer Microsoft 365 environments without any interactive user sign-in, and attempted sign-ins against more than 2,250 downstream tenants. Microsoft was careful to note that none of this came from a flaw in its own systems. The weakness was the trust relationship between a managed service provider and its customers, which is exactly where supply chain attacks live.

The ConnectWise case was not the only one. The FBI alert ties ARTEX to the theft of personal data belonging to roughly 67,000 people from South Korean financial institutions, and lists targets across government, technology, defense, semiconductors, energy, healthcare, manufacturing, telecommunications, and legal services. The Bureau's blunt framing is that the framework lets low-skill actors run multiple complex operations at once. That is the part that should worry defenders more than any single campaign. The skill floor for running a competent multi-stage intrusion just dropped.

Remember that this is the same tool a hobbyist could download from GitHub in July. That is the whole problem.

What it looks like from the defender's chair

Sophos, which has watched ARTEX activity across its customer telemetry, describes something oddly familiar. The attack techniques themselves are not novel: web shells, SQL Server command execution, credential theft, lateral movement. What changed is the character of it. Sophos saw rapid command sequences, the same action retried with small variations, and hundreds of protection events clustered where a human operator would have generated a handful. The tool is noisier and more persistent than a person, because it has no reason to be patient and no fear of tripping an alarm.

That noise is a gift, if you are listening for it. The host-level indicators Sophos published are concrete: temporary files whose names begin with artx, often under C:\Windows\Temp, scratch files following a pattern like i12345_b2.ps1, and command lines carrying markers such as MARKER123 or ARTX123. On the cloud side, the tells are service principal credentials added or roles assigned by an application rather than a person, and application sign-ins coming from a scripted user agent. A burst of nearly identical commands writing output to temp files is not how tired analysts work. It is how an agent works.

The defenses are old. That is the good news.

It would be easy to read all this as a new category of threat that demands new tools. It mostly is not. Read the defensive guidance from the FBI, Microsoft, and Sophos side by side and you get a list that any seasoned team already knows: inventory everything facing the internet, including the partner and supplier applications nobody remembers installing; apply least privilege to application permissions and consent grants; review your Entra audit logs for new credentials, owners, and role changes going back to at least September; stop returning stack traces and SQL errors to anonymous visitors; scan your code for embedded secrets and rotate the ones you find; put rate limits on the APIs that can hurt you.

None of that is a reaction to agentic AI. All of it was good advice last year. What ARTEX changes is the penalty for having skipped it. A misconfiguration that a human attacker might have found in a week, if they bothered, now gets found in an afternoon by a tool that never bothers to get bored. The window between an exposed weakness and an exploited one is closing, and it is closing because the attacker's labor got cheap. A control that was merely advisable when exploitation took effort becomes load-bearing when exploitation is a background process.

What I would check first is the Entra audit log, specifically for application credentials and role assignments added since September, because that is the quiet step where OAuth abuse turns a single stolen secret into tenant-wide access. After that, I would register my own red team's agentic tooling and infrastructure with the SOC, so that the next burst of machine-speed activity on the network can be ruled in or out in minutes instead of becoming the thing the Monday meeting is about. The tools that defend you and the tools that attack you now look the same on the wire. Knowing which is which is becoming part of the job.

References

Gourley, B. (2026, November 14). FBI FLASH on ARTEX: Agentic AI hacking frameworks move from pen testing tool to mainstream threat. OODA Loop. https://oodaloop.com/analysis/ooda-original/fbi-flash-on-artex-agentic-ai-hacking-frameworks-move-from-pen-testing-tool-to-mainstream-threat/

Microsoft Threat Intelligence. (2026, October 21). Storm-0940 uses ARTEX agentic AI in ConnectWise supply chain compromise. Microsoft Security Blog. https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/

Sophos. (2026). ARTEX: Attackers apply agentic AI. Sophos News. https://www.sophos.com/en-us/blog/artex-attackers-apply-agentic-ai