Washington Puts Artex on the Record: Cheap Models, Breaches in Under Three Hours
An advisory from the Federal Bureau of Investigation, published on Nov. 13, 2026, turns vendor inference into an official record of automated intrusions across banks, governments and industry. The price of the models decides who can field one, and the document names no sponsor.
On Nov. 13, 2026, the Federal Bureau of Investigation (FBI) published an advisory that moved Artex out of vendor inference and into the official record. The document, titled Threat Overview of Artex Autonomous Penetration Testing Framework, describes an open-source framework driven by large language models (LLMs) that automates the phases of an intrusion, from reconnaissance to data exfiltration, written in Go and run by a set of agents that need a human only to state the objective. Since at least August 2026, different malicious actors have used it in campaigns against organizations in financial services, government, technology, defense, semiconductors, energy, healthcare, manufacturing, telecommunications and law. In some of those intrusions, the time between selecting a target and holding privileged access stayed under three hours.
Artex arrived with a provenance that reads like ordinary software. It was published on GitHub on July 26, 2026, by a developer using the handle “Autumn-27” under an open-source license, and reached an estimated 1,500 to 2,000 stars in 10 weeks. On Oct. 8, the day after CrowdStrike published its account of malicious use, the author posted a statement denying involvement in the attacks and took the repository down. Censys, a firm that measures internet-exposed systems, notes that the author is also the creator of ScopeSentry, a widely used asset-discovery tool, and that copies survived through archives, translations and rebrands such as ScopeWeaver. By Nov. 4, 2026, Censys counted 1,465 hosts running an exposed Artex console; about 650 of them sat in two identically configured hosting blocks, which leaves roughly 815 distinct-looking deployments, and 155 hosts also ran at least one other piece of attack software. The advisory reports more than 1,600 publicly exposed console instances in early November, while Censys counts hosts, not instances, which explains part of the gap between the two figures.
The advisory’s most consequential passage concerns money. It sorts the model providers the framework plugs into: Anthropic’s Claude and OpenAI’s Generative Pre-trained Transformer (GPT) in one group, DeepSeek, General Language Model (GLM), Kimi and xAI’s Grok in a second, described as open-source or Chinese-origin. Then it reports what the attacker-controlled instances actually ran.
“Chinese models are most prevalent, with DeepSeek Flash family models dominating usage due to their balance of capabilities, relatively low cost, and lack of observable guardrails when framed for penetration testing.”
Two details in that sentence repay a closer read. The reason the document gives is economic: capability at a low price, and no guardrails to trip over when the work is framed as penetration testing. The list itself carries a flaw: Grok is xAI’s model, xAI is an American company, and the advisory’s own network indicators include api.x.ai among the endpoints defenders should watch, so an American model sits inside a group the document labels Chinese-origin.
The advisory describes techniques and lists indicators of compromise. It names no actor, and the attribution comes from private firms with different levels of confidence. Microsoft assessed on Oct. 21 that Storm-0940, a group it describes as aligned with the Chinese state, was responsible for the compromise of ConnectWise Cloud Backup, a multi-tenant application, and recorded the same under-three-hour pattern in automated attacks on hundreds of organizations that began in September 2026.
CrowdStrike described in a report published on Oct. 7, 2026, a separate campaign against South Korean financial institutions that ran from late September into early October. It attributes the activity to an actor it has not named, likely a Chinese speaker and financially motivated, at moderate confidence, and found on a server tied to the campaign an open directory holding Claude Code session histories, Artex configuration files and a Chinese-language pentesting prompt.
Amazon produced its account from decoys. Between Oct. 3 and Nov. 5, 2026, sensors in its MadPot network, configured to look like LLM inference services, logged requests from Artex agents that carried the agents’ own operating memory of their campaigns. Among them was a credential-theft operation built to finance itself with stolen keys and an attempt to clear know-your-customer checks at cryptocurrency exchanges through real-time deepfake injection. Amazon assessed disparate groups with Chinese nexuses, at moderate confidence.
Read together, the four documents keep apart what early reporting merged: one tool, several operators, and a chain of custody for each claim. Nothing published so far links Storm-0940, named in the ConnectWise case, to the actor behind the Korean breaches, and the advisory’s scope covers techniques rather than a sponsor.
The most quoted number in the record is also the softest. Microsoft wrote that Storm-0940 used the framework to “attempt to sign-in to more than 2,250 tenants” after the unauthorized access to ConnectWise; the advisory describes the same episode as a compromise that enabled access to more than 2,250 downstream tenants. TechCrunch reported access to the Microsoft 365 environments of at least 2,250 customers, and ConnectWise told the publication that it does not have the visibility to confirm that number. Attempts, access enabled and access confirmed are three different claims, and only the first carries telemetry. The same story holds a second divergence: the advisory puts the Korean victims at roughly 67,000 people; The Korea Times put the figure at more than 67,000; and TechCrunch reported at least 68,000.
The numbers Microsoft published in detail measure a different offensive. Between Oct. 10 and Oct. 14, 2026, infrastructure at 144.172.64[.]99 sprayed passwords across 327 accounts in eight organizations. Telemetry logged more than 151,000 failed logons, peaking on Oct. 11 at 2,900 failures per minute and concentrating more than 111,000 in a single day. 48 accounts absorbed more than 1,000 failures each, lockouts accounted for nearly 98% of the traffic, and no successful authentication, token issuance or directory change followed. The spray produced no login. The access that counted came through application credentials that already carried permission.
In South Korea, the damage was counted in people. The Korea Times reported on Oct. 5, 2026, that seven financial firms had been hit within days: Shinhan, KB Kookmin, Hana, Busan, Yegaram Savings, Welcome Savings and Hyundai Capital, with more than 67,000 people affected and leaked fields that included resident registration numbers, annual income and loan limits. Three of the banks had spent nearly ₩124 billion, or $92 million, on information security the previous year. President Lee Jae-myung ordered an investigation, and the Financial Services Commission (FSC) convened an emergency meeting the same day.
Japan followed. The security unit of GMO Internet Group (GMO) found six devices dedicated to attacking Japanese companies, Kyodo reported in a dispatch carried by Mainichi Shimbun on Oct. 20, 2026, and, according to the same report, eight firms disclosed leaks in September and October across retail, logistics and finance. Toshiharu Furukawa, the minister in charge of cybersecurity, said the government was analyzing whether there had been malicious use of artificial intelligence.
The clearest first-person account came from a victim. Barracuda’s chief information security officer described a 44-hour intrusion that touched about 10 cloud accounts between Nov. 1 and Nov. 2, 2026, starting from a valid access key that Barracuda had issued to a third-party vendor working in artificial intelligence and never retired after the project ended. The attacker labeled resources with a letter and digits, read secret stores, decrypted infrastructure state files and searched for cryptocurrency wallets; the sequence matched Artex.
Qualys, which read the framework’s source code, frames the shift in units of time. Artex introduces no new class of vulnerability: it finds, chains and retries flaws that were already there, and in its own example the first step is a missing patch and the second is a reused credential. Qualys’s customers now fix critical vulnerabilities in a median of seven days, down from nine a year earlier, against the under three hours recorded in the advisory. The code waits 800 milliseconds between assigning one agent step and assigning the next. Qualys also identifies the commercial edition, Artex Pro, as part of a bundle that ThreatBook sells as RedAgents, a claim Qualys makes without having seen the code of that bundle.
The advisory’s recommendations read like a list of chores: inventory the peripheral systems, such as partner and supplier portals, mobile backends and legacy services, that receive less attention than flagship applications; restrict publishing and administration interfaces; review every tenant where an affected ConnectWise application was installed or authorized, searching by the application’s identifier; examine successful and failed service-principal sign-ins separately.
What the record shows is a market with three inputs. The framework is free, the model is rented by the token, and the key was still valid when the intruder found it. The component that sets the speed of an intrusion is the one with a price list attached. The FBI gave that arithmetic a date and a name, and defenders now hold both the supplier list and the timetable. The inventory that closes the shortest paths remains ordinary work; what changed is how little time the work allows.