Safety#Multi-agent#Cloud

FBI flash alert: ARTEX reached privileged access in under three hours

The FBI's Nov 13 alert says ARTEX, an open-source agentic pentest framework, took some intrusions from targeting to privileged access in under three hours.

An aisle between data center racks, blue status lights on black cabinets, another row of machines at the far end

The FBI’s Internet Crime Complaint Center published flash advisory FLASH-20261113-001 on November 13, and its subject is ARTEX, an open-source, LLM-driven penetration-testing framework that the bureau says criminal groups have been running against financial, government, technology, defense, semiconductor, energy, healthcare, manufacturing, telecom and legal organizations since at least August 2026. The line that traveled furthest is the bureau’s own: some intrusions progressed “from initial targeting to privileged access in under three hours.” The document is 13 pages, marked TLP:CLEAR, and ships with an indicator list.

Key points

  • On the South Korean financial campaign, the alert puts the data loss at “approximately 67,000 individuals.” TechCrunch’s November 16 story says “at least 68,000.” The gap is about a thousand people, and the advisory does not explain the difference.
  • On ConnectWise Cloud Backup, the FBI writes that the compromise “enabled access to more than 2,250 downstream tenants.” Microsoft’s October 21 report says Storm-0940 “used ARTEX to attempt to sign-in to more than 2,250 tenants” and that “the majority of the targeting wasn’t successful.” One sentence later, attempted logins have become access.
  • The tool outlived its takedown. The original repository went offline on October 8, 2026; the FBI says ARTEX keeps circulating through forks, localized variants, archived copies and a commercial edition called ARTEX PRO, with more than 1,600 publicly exposed console instances counted by early November.

Background

ARTEX is written in Go and built around multiple agents. An operator describes a tactical goal in natural language, and the framework’s “super intelligence” agents turn it into machine commands, so the human contribution is mostly the opening objective. It does not train its own model. It calls out to Anthropic’s Claude, OpenAI’s GPT, and open or Chinese-origin models including DeepSeek, GLM, Grok and Kimi. In attacker-controlled instances, the advisory says, Chinese models are the most common, with the DeepSeek Flash family dominant because of capability, price, and, in the FBI’s phrasing, a lack of observable guardrails once the work is framed as penetration testing.

The timeline is short. Developer “Autumn-27” released it on GitHub in July 2026, and it picked up an estimated 1,500 to 2,000 stars in ten weeks. The developer said the tool was meant for authorized testing inside agreed scopes and made the repository unavailable on October 8. We covered Microsoft’s attribution on October 22, when the story was how a tenant admin should audit OAuth grants. This one is about what the FBI chose to put in a national advisory, and what happened to the numbers on the way out.

What the alert says

The same fact carries different figures depending on which document you read:

ClaimFigureSource
South Korean customer dataabout 67,000 peopleFBI alert (2026-11-13)
Same breachat least 68,000 peopleTechCrunch (2026-11-16)
ConnectWise downstream tenants“access to” more than 2,250FBI alert
Same tenants“attempt to sign-in” to more than 2,250, mostly unsuccessfulMicrosoft (2026-10-21)
Exposed ARTEX consolesmore than 1,600FBI alert, citing Censys
Same finding1,465 hosts as of November 4Censys (2026-11-05)
GitHub stars1,500 to 2,000 in ten weeksFBI alert
Intrusion speedinitial targeting to privileged access in under 3 hoursFBI alert

The two console rows need separating. Censys counted 1,465 hosts, roughly 650 of them inside two hosting blocks with identical configurations, which leaves about 815 distinct-looking deployments; the FBI rounds upward to “more than 1,600.” That work is exposure research, not a breach report, and our November 10 digest listed it as a candidate it passed on. Both numbers can be right, but a reader citing one should know which measurement it came from.

Most of the advisory is a table mapping observed activity to MITRE ATT&CK phases, from reconnaissance through command and control. The techniques themselves are not worth repeating here. Two observations in it are useful to defenders. First, the FBI says these agents are good at finding peripheral systems: partner and supplier portals, mobile backends, employee apps, sales-support tools, legacy services, third-party integrations, applications built inside individual business units. Those rarely get the security attention the flagship product gets. Second, the traffic is noisy and will trip alerts, but the speed leaves too little time for a coordinated response before the objective is complete.

The indicator list contains one category that is still uncommon in advisories: outbound calls to LLM APIs, with api.anthropic.com, api.deepseek.com, open.bigmodel.cn and api.x.ai named, plus traces of self-hosted LLM gateways and reseller software such as CLIProxyAPI and New API. That is more useful to a defender than another blocklist entry. The FBI also warns that its IPs sit on cloud and virtualized infrastructure where addresses get reassigned, so the list describes historically observed infrastructure and should be checked against current telemetry before anyone blocks it.

What others say

Microsoft’s October 21 report attributed the ConnectWise activity to Storm-0940, which it calls “a China-state aligned actor,” and recorded “multiple instances where the time from initial targeting to privileged access was less than 3 hours.” In that report, 2,250 is the number of tenants the group tried to sign in to, followed by the note that most of the targeting failed. The FBI’s summary turns that into access, and TechCrunch describes “at least 2,250 customers.” Three retellings, one verb change, three different levels of severity.

CrowdStrike describes a different cluster. In an October 30 report covered by CyberScoop, the firm tracks a China-linked group it calls Vapor Panda, which used ARTEX and DeepSeek against more than 200 organizations in at least 14 countries from October 17, with about 140 sensitive targets compromised in 11 days and roughly 70% assessed as confirmed or moderate-confidence intrusions. CrowdStrike put the tempo in comparable terms: about 40 commands an hour for this kind of operation historically, up to 1,500 an hour with ARTEX, and requests as often as every 10 seconds. It counted about 45 agentic attacks in September and 126 in October. A CrowdStrike official told CyberScoop this is the first time the firm has seen a nation-state actor operate at that scale with agents.

The South Korean campaign carries a third name. Recorded Future tracks it as TAG-217; the FBI advisory gives no actor name at all and cites CrowdStrike and Zscaler analysis instead. Three names, two clusters, and no public report that settles whether they are the same people. What is settled is that one framework was used by different groups inside a single quarter.

What defenders can do

The FBI’s recommendations sort into assets, identity, applications and evidence. The parts most teams can act on:

  1. Count peripheral systems when inventorying internet-facing assets. Partner and supplier portals, mobile backends, employee apps, sales-support tools, legacy services and business-unit applications are the entry points the advisory names.
  2. Move management planes off the public internet. Cloud consoles, Kudu/SCM publishing endpoints, network device management, hypervisors and backup platforms belong on trusted networks or dedicated admin paths, and basic publishing authentication should be off where nothing needs it.
  3. Search every Microsoft 365 tenant that has or had a ConnectWise application by application ID, not display name, including integrations nobody uses anymore. Review successful and failed service-principal sign-ins separately rather than checking only the published IP ranges, and read Entra audit records for new credentials, new applications and service principals, app-role assignments and consent grants. Start the review at October 1, 2026.
  4. Put rate limits, recipient caps and per-session record caps on high-impact business APIs, and alert on sequential lookups. Keep stack traces, SQL errors and internal paths away from unauthenticated users.
  5. Preserve before you delete. Entra sign-in and audit records, Purview, SIEM data and Graph telemetry should be captured first; revoke credentials and grants only after that. Keep ConnectWise connections disabled until the vendor publishes restoration instructions, and do not delete enterprise applications as a first move.

Our take

ARTEX is not a vulnerability. It is a layer of automation on top of things that were already wrong: peripheral systems answering on the public internet, application credentials that never expire, trust granted to third parties and never reviewed. There is no patch for it. What can change is the asset inventory, the credential lifecycle and the trust boundary. Treating this as an “AI attack” problem tends to buy the wrong tools; treating it as someone chaining old problems together in a few hours points at work most teams already know they owe.

The three-hour figure matters less for its precision than for what it does to process. An alert has to be read, understood, confirmed across teams, and turned into a decision about which service to cut. That loop was already measured in hours. Once the first half of an intrusion runs in minutes, the bottleneck becomes who is allowed to decide. The pattern of credentials and trust failing first while agents spread the consequences is one we laid out in what recent agent incidents had in common; this is the same pattern with a faster execution layer.

The gaps are worth naming. Nobody has published how many of those 2,250 tenants were actually entered. There is no public account of victim notification or regulatory filing in South Korea. The FBI did not name the actor behind the Korean campaign. And no independent reporting yet establishes who buys ARTEX PRO or where it is sold. TechCrunch reports Microsoft’s assessment that the intruders are “likely Chinese,” which is an assessment, not a finding.

If you do one thing, pull the internet-facing asset list and the Microsoft 365 application consent list, and check them by application ID. None of that involves AI, but it is the only part of this story a defender can finish alone.