# One goal, then the agents did the rest: what the FBI's ARTEX warning means for your team

<!--
title: "One goal, then the agents did the rest: what the FBI's ARTEX warning means for your team · Cork"
description: "The FBI's FLASH on ARTEX describes AI agents that take a goal in plain language and carry an intrusion from scouting to stolen credentials, some reaching privileged access in under three hours. When an attack finishes that fast, an alert is too late; the controls that work are the ones already in the path."
canonical: "https://corkcli.com/blog/fbi-artex-warning-ai-agent-intrusions"
last_updated: "2026-11-16"
published: "2026-11-16"
author: "Kevin Vera"
-->

The FBI's FLASH on ARTEX describes AI agents that take a goal in plain language and carry an intrusion from scouting to stolen credentials, some reaching privileged access in under three hours. When an attack finishes that fast, an alert is too late; the controls that work are the ones already in the path.

By Kevin Vera · November 16, 2026 · [News](https://corkcli.com/blog/c/news) · 5 min read

On 13 November 2026 the FBI published a FLASH, FLASH-20261113-001, on ARTEX: "an open-source, large language model (LLM)-driven agentic platform that threat actors have weaponized to automate multiple phases of cyber intrusions at machine speed, from reconnaissance through data exfiltration." The person running it sets a goal in plain language. The framework, in the FBI's words, "translates high-level tactical goals described in natural language into actionable machine commands executed by specialized super intelligence (SI) agents, requiring minimal human interaction beyond initial objective setting." Some intrusions went "from initial targeting to privileged access in under three hours."

Three hours is the number that matters for anyone deciding how to defend. Most security is built around noticing something and then responding. When the whole attack can be over before the first person looks at the first alert, the defences that count are the ones already standing in the way.

## What the FLASH says ARTEX did

- **Took a goal and ran with it.** Check Point Research, which the FLASH cites, looked at two campaigns and found no human steering at all: "Beyond the initial task goals, the agents ran fully autonomously." One real goal it quotes: "get a p12 certificate + password".
- **Scouted.** The FLASH lists "fuzzing web directories across 500+ requests" and "WHOIS asset mapping and active vulnerability scanning".
- **Stole credentials.** Kerberoasting for service tickets, DCSync to take the domain's krbtgt key, reading LAPS passwords, and "stolen VPN configurations and leaked credentials".
- **Tried to move.** The one lateral move the FLASH spells out, toward domain controllers with Pass-the-Ticket, WinRM, Kerberos relay and Pass-the-Hash, was "blocked by endpoint protection". Microsoft, whose October report is the source of the three-hour figure, describes agents pivoting into customer tenants with stolen application credentials instead.
- **Got far, sometimes.** The three-hour figure comes from Microsoft's report on Storm-0940, which it calls "a China-state aligned actor": "multiple instances where the time from initial targeting to privileged access was less than 3 hours", in organisations' cloud infrastructure. Microsoft also notes "the majority of the targeting wasn't successful."

The damage named in the FLASH: personal information on about 67,000 people, and a supply-chain compromise of ConnectWise Cloud Backup "that enabled access to more than 2,250 downstream tenants." As of early November, researchers had found "over 1,600 publicly exposed ARTEX console instances." The tool was first published on GitHub in July 2026; its developer took the repository down on 8 October, and copies are still in use. Several groups have used it "since at least August 2026."

## Why alarms aren't enough

None of the FLASH's mitigations is "watch more closely." They remove paths before anyone has to react:

- "Inventory all internet-facing applications, including peripheral systems."
- "Apply least privilege to multi-tenant application permissions and consent grants."
- "Put guardrails around high-impact business APIs: rate limits, recipient limits, transaction thresholds."
- "Move enterprise applications behind Zero Trust access solutions."
- Scan "CI/CD pipelines, databases for embedded secrets", because leaked credentials were one of the ways in.

An agent that works at machine speed doesn't wait for a ticket to be triaged. A rate limit, a scoped token or a transaction threshold stops it at the moment it acts, whether or not anyone is watching.

## The same property, on your side

Your own agents work at machine speed too. Most teams running Claude Code, Codex or Cursor on a real codebase have handed those agents shell access, repository write access, and often cloud credentials. ARTEX is a reminder of what an agent can get through in an afternoon. The design questions are the same ones the FBI is answering for defenders: what is the agent allowed to touch, what does it have to ask for, and what stops it at the moment it acts, rather than after?

One line in the FLASH stood out to us. Among the evidence the FBI asks victims to report is "evidence of multiple sub-agents overwriting each other or generating duplicate requests." Even an attacker's agents run into the problem every team running parallel agents has: several agents acting at once, each unaware of what the others just did.

That is the problem Cork works on, for your own coding agents. Cork is not a security product, and it does nothing to stop an attacker. But its controls sit in the same place the FBI's advice points to: in the path, before the action lands. A write against code another agent just changed is held, and the agent is told what changed. A handoff that would delete files or data, rewrite git history, tear down infrastructure, deploy to production, or change secrets or access waits for a person's yes, whoever sent it.

## What to do this week

1. Work through the FLASH's own list: an inventory of internet-facing apps, least privilege on application consent grants, limits on your high-impact APIs, and a secrets scan of CI/CD and your databases.
2. Give your own agents the least privilege they can work with: scoped tokens, and no production credentials on a laptop that runs agents unattended.
3. Make destructive actions ask: deleting data, rewriting history, deploying to production, touching secrets.
4. Keep logs where the agent can't reach them. In the [Strix case](https://corkcli.com/blog/strix-ai-agent-cloudsek-silent-lynx) CloudSEK describes this week, cleaning up logs was one of the agent's jobs.

## Sources

- FBI FLASH-20261113-001, "Threat Overview of ARTEX Autonomous Penetration Testing Framework", 13 November 2026 (TLP:CLEAR): https://www.ic3.gov/CSA/2026/261113.pdf
- Microsoft Threat Intelligence, 21 October 2026: https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/
- Check Point Research, "Hacking on autopilot: inside ARTEX cybercrime campaigns", 5 November 2026: https://research.checkpoint.com/2026/hacking-on-autopilot-inside-artex-cybercrime-campaigns/

## Frequently asked questions

### What is ARTEX?

An open-source framework that runs AI agents through the stages of an intrusion, from reconnaissance to data exfiltration, from a goal written in plain language. It was published on GitHub in July 2026 and taken down in October; the FBI says several groups have used it since at least August 2026.

### Did ARTEX agents get admin access in under three hours?

The FBI says "privileged access", and the source is Microsoft's report on one actor, Storm-0940: in multiple instances, in organisations' cloud infrastructure, it took less than three hours from first targeting. Microsoft adds that most of the targeting wasn't successful.

### Does Cork protect against ARTEX?

No. Cork coordinates your own coding agents on your codebase; it is not a security product. What it shares with the FBI's advice is where its controls sit: in the path, before a write lands or a destructive handoff runs.

## Related reading

- [The attacker kept office hours, the agent did the work: inside CloudSEK's Strix report](https://corkcli.com/blog/strix-ai-agent-cloudsek-silent-lynx): CloudSEK found an attacker's server open on the internet, and on it an AI agent named Strix: tasked over Telegram, given a persona file and pre-approved destructive commands, with a tunnel into Tajikistan's national railway network for nearly two months. Its building blocks are the ones teams use for their own agents.
- [From one machine to a team: moving from Cork Free to agent hours](https://corkcli.com/blog/cork-free-to-team): Move from Cork Free to agent hours: what changes when agents on teammates' machines see each other, how cork login and cork upgrade work, and what it costs.
- [Coordination layer for AI coding agents: what it is and what it does](https://corkcli.com/blog/coordination-layer-ai-coding-agents): A coordination layer for AI coding agents sits under the tools that run agents and keeps parallel agents aware of each other's changes so the work fits.
- [Landing: rebase, test and push in one step for AI coding agents](https://corkcli.com/blog/landing-rebase-test-push): Landing means getting finished work onto the shared branch in one step: rebase onto everyone else's commits, rerun the tests, then push only if they pass.

[All articles](https://corkcli.com/blog.md) · [Cork in Markdown](https://corkcli.com/index.md) · [llms.txt](https://corkcli.com/llms.txt)

This page in HTML: https://corkcli.com/blog/fbi-artex-warning-ai-agent-intrusions
