Skip to content
AnalysisAI

ARTEX: The AI Hacking Tool That Outlived Its Own Deletion

The FBI describes ARTEX intrusions reaching privileged access in under three hours. Researchers are also tracing Hermes, a legitimate AI assistant that criminals have turned into an operator’s console.

Two FBI cyber agents study computer monitors in a darkened Jacksonville office; one leans over the other’s workstation.
FBI cyber agents in Jacksonville, Florida, in an agency file photograph dated August 31, 2012. The bureau’s November 13, 2026 advisory describes malicious use of ARTEX.FBI · FBI publicity-use permission

Some intrusions using ARTEX reached privileged computer access in under three hours, according to the FBI. By the time the bureau issued its November 13 warning, the AI-assisted hacking framework’s original developer had already removed its public repository. Deleting the code’s first home had not stopped its use across finance, government and technology.

The code survived in copies and continuations. So did the ability to connect it to language models and let agents carry out tasks with actual tools. ARTEX’s spread has become a problem for companies whose credentials and trusted applications give those agents somewhere to go. It has also become a problem for people downloading the software: researchers found a prebuilt copy containing a back door of its own.

View FBI Dallas's post on X

What the software does

ARTEX is a security-testing program written in the Go programming language that connects to external AI language models. A planner divides a goal into tasks, agents use tools, and the system records what they tried and found. It is not a foundation model like the underlying GPT or DeepSeek systems. The same broad arrangement can support authorized attempts to find security weaknesses, known as penetration testing, or intrusions conducted without permission.

The original developer used the public handle Autumn-27. The project appeared in July and its original repository disappeared on October 8, after reports of abuse. Removing that repository did not revoke code people had already obtained. Independent continuations remain publicly available, with their own maintainers and changes.

The 2,250 figure counts attempts

Microsoft attributes the ConnectWise Cloud Backup campaign to Storm-0940, which it assesses as aligned with the Chinese state. Its October 21 report describes stolen credentials for applications already authorized inside customers’ Microsoft 365 environments. That permission gave the attacker a route through a trusted application. Microsoft reported no exploitation of a Microsoft product vulnerability; how the credentials were stolen remained undisclosed.

An earlier password-spraying phase, testing passwords across accounts, generated more than 151,000 failed sign-ins against 327 accounts in eight organizations. Microsoft found no successful authentication in that phase. Later activity attempted sign-ins against more than 2,250 customer environments, or tenants. That is not 2,250 breached companies; authentication alone would not show what data was taken.

This Cloud Backup case is separate from November’s ScreenConnect incident. Sharing a vendor does not make them the same attack.

The cases do not all belong to one actor

CrowdStrike’s October 7 South Korean finance investigation described an unknown actor using ARTEX. The FBI’s later advisory cited approximately 67,000 affected individuals in the Korean cases. TechCrunch reported at least 68,000; the published counts differ. CrowdStrike separately attributed a different campaign to the China-nexus group it calls Vapor Panda: more than 200 organizations targeted in at least 14 countries over 11 days, with confirmed post-exploitation activity at nine organizations. Targeting 200 organizations is not evidence that all 200 were breached.

WorkOS supplies a different and uncomfortable account. In November, founder and CEO Michael Grinich wrote that “a security researcher contacted us and claimed responsibility.” WorkOS says the person provided evidence that it and an independent forensic investigator matched to company logs. The researcher said an ARTEX test escaped an isolated environment and interacted with WorkOS unintentionally.

View Michael Grinich's post on X

The affected system was an internal sales and customer-communications application, rather than WorkOS’s production authentication platform. WorkOS described a legacy internet-facing system, a predictable development credential and gaps in limits on repeated requests and in activity monitoring. The company corroborated the researcher’s responsibility; the claimed accident, deletion of data and unexplained escape mechanism remain attributed accounts. They do not establish that an AI independently decided to rebel.

CJ Moses, writing Amazon’s November 13 investigation, described collecting agents’ operating memory through decoy service interfaces. Amazon attributed the activity to ARTEX with moderate confidence. Those records offered a view of credential hunting, but the company said it could not independently confirm activity against other organizations named in that memory. An agent’s account of what it did needs corroboration too.

CJ Moses wears a navy blazer and light blue shirt while standing with his hands on a chair in a sunlit meeting room.
CJ Moses, the author of Amazon’s November 13 ARTEX investigation, in an undated Amazon file photograph. Amazon attributed the activity it observed to ARTEX with moderate confidence.Amazon / AWS · Amazon editorial press permission

Which models are doing the work?

The framework’s external-model design lets different operators choose different services. CrowdStrike’s Korean report identified DeepSeek v4.1-flash as the main model, with GLM 5.3 and Grok 4.6 in additional Claude Code sessions. The name of that client does not prove that Claude supplied the underlying model; a proxy can route requests elsewhere.

Check Point studied ten exposed ARTEX databases and found Chinese models prominent in that sample. One operator’s two consoles consumed about 36 billion tokens from October 14 to 29, mostly cached input, meaning material reused in requests to the model. Check Point estimated a historical cost of $700 to $1,500. Those are the researchers’ sample and estimate, not an ARTEX-wide cost or a current provider quotation. No rigorous model-by-model ARTEX performance benchmark is established by those observations.

Hermes can carry a campaign between conversations

Hermes Agent is a legitimate, general-purpose assistant from Nous Research, released February 25 under the MIT license. It connects language models to tools, keeps persistent memory and searchable past sessions, runs scheduled jobs and offers a web console. It is separate from Nous’s Hermes language models. Criminal use of the agent does not make its developers participants in those crimes.

View Nous Research's post on X

Its reusable skills are plain files called SKILL.md, which the agent can create and edit. Saving a procedure lets another session use it without reconstructing the whole conversation. That changes the instructions an agent can consult; it does not mean the software has retrained the underlying model’s weights.

On November 17, security researcher Thomas Roccia reported that Storm-3358 was using Hermes alongside ARTEX to compromise organizations. This specific connection rests on his reporting alone. Microsoft’s November 6 account of Storm-3358 describes ARTEX but never mentions Hermes. No separate vendor investigation reviewed for this article corroborates Roccia’s Storm-3358 claim.

View Thomas Roccia's post on X

The distinction matters because other researchers have documented Hermes abuse by different operators. Those investigations establish a broader pattern; they cannot authenticate the contents of a particular criminal group’s computers.

Roccia’s screenshot shows an excerpt from a 27-kilobyte, 288-line file named redteam-detection-surface-and-resilience. Its description concerns the likelihood of detection and whether access could be recovered after an intruder is removed. The excerpt lists five related skills: attacker-host-opsec-hardening, engagement-reverse-analysis, post-exploitation-foothold-gating, c2-channel-liveness-triage and controlled-access-revalidation.

Only part of one file is visible. Its internal-license label and author field do not independently establish who wrote it. None of those six names appeared in the official repository tree checked for this article, but that limited search cannot prove that no public copy exists elsewhere. The complete files were not available for review.

The important change is where an operator can keep instructions. A criminal can preserve procedures outside a single chat, alongside an agent that remembers earlier work. ARTEX already has its own planning and memory. Hermes can supply another layer across sessions and tools; it is not a prerequisite for every ARTEX attack.

Separate investigations show what the console can do

Gambit Security’s September 22 investigation recovered an operator’s staging server in a retail card-skimming campaign. It reported 121 loaded Hermes skills, including 78 attack skills and one intended to remove safety filters. Strix searched for vulnerabilities and Cairn handled exploitation under Hermes’s direction.

Eyal Sela, Gambit’s director of threat intelligence, wrote: “Hermes is the operator’s console for orchestrating the activity and for direct hacking activities.” His report acknowledged incomplete records and outcomes based partly on logs and agents’ own accounts. Those limits apply to the reported impact, even when the recovered configuration is concrete.

View Eyal Sela's post on X

Unit 42 documented another operator on July 30, using Hermes with DeepSeek and sending tasks over Telegram. Its account describes exploitation attempts against more than 460 internet-facing targets, combining automated and manual activity. That number counts attempted targets, not 460 breaches. The investigators described a Chinese-speaking operator, which does not establish nationality or Chinese government direction.

Elastic Security Labs’ November 13 study examined five offensive frameworks and found none enforced authorization in code by default. It cited Gambit’s Hermes finding when discussing the wider ecosystem. Citing that report does not turn Elastic into a second investigator of the same campaign, or a corroborating source for the separate Storm-3358 allegation.

View Elastic Security Labs's post on X

What ARTEX and Hermes each contribute

SoftwareRoleEvidence limit
ARTEXPlans security tasks, runs tools and records findings.Targeting and attempted sign-ins are not automatically confirmed breaches.
Hermes AgentGeneral-purpose memory, skills, sessions and orchestration; it can also act directly.Separate documented abuse cases do not corroborate every proposed actor link.

The specific Storm-3358–Hermes link is Roccia’s single-source report. The other investigations describe different campaigns.

FBI advisory; Nous Research; Gambit Security; Unit 42; Thomas Roccia.

For defenders, the two layers point to different evidence: the tools taking actions, and the sessions and skills directing them. Elastic argues for detecting behavior, timing and sequences of actions rather than relying on a framework’s name or network address. Renaming a program can defeat a name-based rule while leaving its account access and requests observable.

Four masked participants sit around a table during an FBI cybersecurity workshop in Charlotte; one gestures while the others listen.
FBI Special Agent Kathryn Sherman leads a workshop for cybersecurity executives in Charlotte, North Carolina, in October 2021. File photograph.FBI · FBI publicity-use permission

The copy that put its own users at risk

BadByte’s investigation found a malicious prebuilt ARTEX software package, distributed as a container image, with a command-execution implant accessible before authentication. The change was absent from the public source code. Someone inspecting that source alone would therefore have missed the compromised binary they actually ran.

Two malicious image namespaces recorded 7,946 combined pulls by November 3. Pulls count downloads, not unique people or successful compromises. BadByte’s November 5 update says Docker removed both images. The container ran as root, but that fact alone does not show an escape into the host computer. The report documented a poisoned build distributed through copies of the project, not a fake download website.

Where the surviving versions live

Public GitHub continuations include BaaRai/ARTEX, RuoJi6/ARTEX and Hinln/ARTEX; their links are in Sources. RuoJi6’s verified release snapshot included version 0.3.30, published November 14. A repository name or a familiar-looking interface does not identify which binary someone obtained.

Cloudflare reports roughly 12,000 repositories with ARTEX lineage and more than thirtyfold growth in its detected daily requests since October 1. Lineage, forks, public consoles and victims are different populations. The observations show continued proliferation around the original removal; they do not prove that deleting the repository caused the growth.

What would a ban actually remove?

The policy fight has already begun. In November 12 reporting, TechTarget described Microsoft global CISO Igor Tsyganskiy urging worldwide regulation and warning that ARTEX was “spreading like fire.” Roccia told the publication: “Criminals won’t respect a ban.” Fresenius CISO Boban Kršić argued that “the people who actually get hurt are our own red teams.”

Hermes makes that dispute harder to reduce to one prohibited product. Memory, scheduled jobs and reusable instructions also serve ordinary assistants. A restriction defined around those features would reach legitimate software too. Policymakers would have to identify the conduct or capabilities they meant to restrict, while companies would still need to revoke stolen access and investigate what actually happened.

The same trust problem appears in the Optimizely supply-chain attack: organizations inherit risk through systems they already use. Google’s exploit_agent report examines a separate attempt to automate vulnerability discovery and exploitation.

What we expect next

We expect incident responders to recover more general-purpose agent consoles and reusable skills beside offensive frameworks. Separate investigations have already found that combination. Evidence that it remains confined to a few operators would weaken that prediction. The developments to watch are independently verified intrusion outcomes, revoked application permissions and build histories that identify what people actually ran, rather than more headlines converting attempted sign-ins into thousands of breaches.

Common questions

What is ARTEX?

ARTEX is a Go-based framework that connects external language models to agents and security-testing tools. The FBI’s November 13, 2026 advisory describes malicious use of that framework.

Is ARTEX itself a language model?

No. ARTEX coordinates tools and agents while relying on external models; it is not itself a foundation model such as GPT or DeepSeek.

Did ARTEX breach 2,250 ConnectWise customers?

That is not established. The more-than-2,250 figure concerns attempted sign-ins against downstream tenants in the October 2026 Cloud Backup case.

Does the FBI say every ARTEX intrusion takes three hours?

No. Its November 13 advisory describes some intrusions progressing from initial targeting to privileged access in under three hours.

Were all 200 Vapor Panda targets breached?

No such finding was reported. CrowdStrike described more than 200 targeted organizations but confirmed post-exploitation activity at nine in its October 29 report.

What happened at WorkOS?

WorkOS says a researcher supplied evidence matching its logs and claimed an ARTEX test unintentionally reached an internal sales application. The mechanism of the claimed escape remained unexplained.

Which models have investigators observed?

CrowdStrike identified DeepSeek v4.1-flash as primary in its Korean case and GLM 5.3 and Grok 4.6 in additional sessions. Those observations are not an ARTEX-wide benchmark.

Was there a backdoored ARTEX build?

Yes. BadByte reported an implant in prebuilt container images whose malicious changes were absent from the public source. Its November 5 update says Docker removed the two images.

Do 7,946 image pulls mean 7,946 victims?

No. BadByte’s dated November 3 snapshot counted image downloads, not unique users or confirmed compromises.

Where can the surviving ARTEX projects be found?

Public continuations include BaaRai/ARTEX, RuoJi6/ARTEX and Hinln/ARTEX on GitHub. Their existence is not a safety certification for a downloaded build.

What is Hermes Agent?

Nous Research released Hermes Agent February 25, 2026 under MIT licensing. It is a general-purpose agent with memory, reusable skills, past sessions, scheduled jobs and a web console.

Has Microsoft confirmed Storm-3358 used Hermes?

No reviewed Microsoft write-up confirms it. The specific November 17 claim is Thomas Roccia’s reporting; Microsoft’s November 6 account mentions ARTEX, not Hermes.

How many Hermes skills did Gambit report finding?

Gambit’s September 22 investigation described 121 loaded skills, including 78 attack skills, in a separate retail campaign. Some claimed outcomes remained dependent on incomplete logs and agents’ accounts.

Were 460 companies breached in Unit 42’s Hermes case?

That is not what the July 30 report establishes. More than 460 describes attempted targets, not a confirmed breach count.

Are the six Storm-3358 skill files available to inspect?

The reviewed evidence shows an excerpt from one 288-line file and five related names, not six complete files. None of those names appeared in the official repository snapshot checked.

Sources

All news
John Hultquist speaks into a microphone with bookshelves behind him.
Engineering

Google’s Autonomous Cyberattack Warning Came With a Sales Pitch

Michael Kratsios, seated in a white armchair, holds a microphone and gestures while speaking onstage at CES 2026 beside Kinsey Fabrizio.
AI

Name and Shame for Chinese AI: Trump Officials Weigh a Disclosure Plan

A Waymo Jaguar I-Pace with roof-mounted sensors driving on a city street
Strategy

Choosing an AI Use Case a Team Can Evaluate

Have data to license or need a source?

Tell us what your company has or what your team needs.

Talk to SnowRock