Waxell

Does the FBI ARTEX alert reach your company's AI agents?

Partly. ARTEX is the attacker's AI agent, but the FBI's advice on inventory, privilege and high-impact actions also covers the agents you run.

Part of: AI Agent Security Threats: Prompt Injection and Beyond

Logan Kelly5 min read

Partly. The FBI's ARTEX alert is about attackers using an AI agent framework against you, not about the AI your company runs, but its core advice (know what can reach your systems, watch privilege, and put a check in front of high-impact actions) applies to your own AI agents too. For the wider picture, see our guide to AI security threats.

On 13 November 2026 the FBI published FLASH-20261113-001, "Threat Overview of ARTEX Autonomous Penetration Testing Framework", through IC3. It is marked TLP:CLEAR, which means it can be shared without restriction. Read it here: FBI FLASH-20261113-001 (PDF).

What does the FBI say about ARTEX?

ARTEX is an open-source penetration testing framework made of several agents driven by large language models. Its developer said it was meant for authorized security testing. According to the FBI, malicious actors have used it since at least August 2026 to automate intrusions from reconnaissance through to data exfiltration. Targets span financial services, government, technology, defense, semiconductors, energy, healthcare, manufacturing, telecommunications and legal.

Speed. The FBI says some intrusions went from initial targeting to privileged access in under three hours. Some, not all, but it shrinks the window a security team plans around.

Scale. Citing industry research, the FBI notes that researchers found more than 1,600 publicly exposed ARTEX consoles in early November 2026, and says total deployments are likely much higher.

No model of its own. ARTEX plugs into outside model providers. The original developer removed the GitHub repository on 8 October 2026, after its malicious use became public, but forks, archived copies and a commercial version (ARTEX PRO) remain available.

The alert describes two notable cases: a campaign against South Korean financial institutions that exposed personal data of about 67,000 people, and a supply chain compromise of ConnectWise Cloud Backup customers that led to sign-in attempts against more than 2,250 downstream tenants.

What does the FBI recommend?

The recommendations fall into four groups.

  1. Shrink the external attack surface. Inventory every internet-facing application, including peripheral and third-party systems such as partner and supplier portals. Restrict admin and publishing interfaces to trusted networks or dedicated admin paths. Move enterprise apps behind Zero Trust access where you can.
  2. Harden identity and access. Apply least privilege to multi-tenant app permissions and consent grants. Monitor credential additions, privilege assignments and unusual app-only activity. Review service principal sign-ins and Microsoft Entra audit records. If you use ConnectWise, review every Microsoft 365 tenant where its affected apps are or were installed or authorized.
  3. Protect applications. Put guardrails around high-impact business APIs, such as rate limits, recipient limits and transaction thresholds, and alert on unusual volumes of records accessed. Do not return detailed error messages to unauthenticated users. Scan for embedded secrets.
  4. Preserve evidence and respond. Preserve sign-in, audit and security logs before irreversible changes, without delaying urgent containment. ConnectWise customers should review related activity from at least 1 October 2026 and keep affected connections disabled until ConnectWise issues restoration instructions. Report incidents to your local FBI field office and to IC3.

What the alert does not say

ARTEX is the attacker's agent, running on the attacker's infrastructure. The alert does not mention AI governance, MCP, AI coding assistants or the AI tools your employees use. It does list some LLM-related warning signs, such as unfamiliar LLM proxies in egress logs and unusual token use or API billing, but as indicators of ARTEX activity. If anyone tells you the FBI recommends their product category, including us, read the PDF again.

Most of the FBI's recommendations sit outside what Waxell does: web application hardening, identity provider reviews, backup platform fixes. Do those first.

Where do your own AI agents come in?

The FBI tells you to inventory what attackers could reach, to watch privilege and app-only activity, and to put guardrails around high-impact actions. In many companies, AI agents and AI tools now hold credentials to exactly those systems: email and documents, the ERP, the CRM, the code repository. An agent that someone connected to Microsoft 365 with admin rights is part of the surface the FBI is describing, whether or not it is on anyone's list.

Attackers running agents that can reach privileged access within hours do not need a new way in if one of your own agents already has the keys.

Three questions worth asking now:

  1. What AI is running on our employees' machines, and what is it connected to?
  2. Which high-impact actions can our agents take without a person approving them?
  3. If a tool one of our agents relies on changes underneath it, will anyone know?

How Waxell helps with those three questions

Finding what is running. Waxell Endpoints finds the AI already running on employee laptops. It finds MCP servers configured in popular AI coding tools and desktop assistants, and flags those that bypass the Gateway.

Putting a person in front of high-impact actions. In the Waxell MCP Gateway, admins can configure human approval holds in three modes: a named reviewer, in-band confirmation, or async. New tenants get default deny rules for permanent deletes, bulk deletes and history rewrite.

Knowing when a tool changes. The MCP Gateway fingerprints tools (name, description, input schema). A deny_drift rule, which an admin sets up, re-checks the tool live at call time and also denies untrusted or unverifiable tools. The Gateway also scans tool descriptions and nested schema descriptions for prompt injection on every fingerprint push.

Keeping the record. The MCP Gateway keeps a tool-call audit log: who called what tool, which decision applied and which rules fired, exportable as CSV.

One limit to keep in view: the MCP Gateway governs the calls that go through it. Agents holding direct upstream credentials, locally registered MCP servers and unconfigured clients bypass it. That is why finding them on the device comes first.

To be plain about scope: Waxell does not detect ARTEX or the attacker's infrastructure. Its job is the AI inside your company: finding AI tools and agents on employee machines, and governing the tool calls that go through the MCP Gateway.

FAQ

ARTEX is an open-source penetration testing framework built from several AI agents driven by large language models. The FBI says malicious actors have used it since at least August 2026 to automate intrusions, from reconnaissance to data exfiltration.

Yes. The original developer removed the GitHub repository on 8 October 2026, but the FBI says community forks, archived copies and a commercial version called ARTEX PRO remain available.

No. The alert covers attackers using ARTEX and how to defend against intrusions. It does not mention AI governance, MCP or AI coding assistants. The link to your own agents is an inference from its advice on inventory, privilege and high-impact actions.

Sources

Bring your security team: book a 30-minute walkthrough.