Microsoft: Chinese-linked group used Artex against more than 2,250 Microsoft 365 tenants
An open-source hacking tool published on GitHub in July of this year is now officially classified as a mainstream criminal weapon. The FBI advisory FLASH-20261113-001 describes how Artex connects large language models to standard penetration-testing tools — and how removing the repository itself turned out to mean very little.
On November 13, 2026, the FBI published a 13-page alert with the classification TLP:CLEAR, numbered FLASH-20261113-001. According to secondary sources covering the advisory, it marks a turning point: a US law enforcement authority is now treating autonomous AI hacking tools not as a research curiosity, but as an active, mainstream threat to infrastructure worldwide. The advisory itself is not available in the source material for this article — all figures and quotations below are relayed through secondary coverage and attributed accordingly.
The core of the warning is speed. "ARTEX substantially reduces the time required to chain together attack techniques, with some intrusions progressing from initial targeting to privileged access in under three hours," the FBI warned, according to the NY Post's rendering of the advisory. The agency also describes the tool as allowing hackers with very little technical skill to carry out complex operations against many targets simultaneously.
What Artex actually is
Artex is not an AI model. It is an orchestration layer written in the Go programming language, published as open source on GitHub in July 2026 by a pseudonymous developer with the handle Autumn-27, according to WIRED's reporting on November 18. Since then, it has spread through several hacking communities.
Technically, the framework works like a conductor that connects external AI models — both frontier models and open-source models, connected via API — to a series of existing penetration-testing tools and processes, WIRED writes with reference to security researchers. The agents break down plans, execute tasks and report findings, driven by natural-language prompts. In other words: the leaps in capability come not from a new model, but from an agent orchestrator binding together models such as ChatGPT and Claude with tools the hackers were already using. This is what the FBI advisory, according to the Hindustan Times' coverage, describes as an "autonomous, artificial intelligence-driven penetration-testing framework" that has been "weaponized" since at least August 2026, targeting banking, the public sector, technology, defense, healthcare and telecom.
AWS's chief security officer CJ Moses summarized the development in a blog post cited by WIRED: "Agentic frameworks have progressed from experimental use to sustained, multi-objective offensive operations that run with minimal human intervention."
Documented attacks — and who is behind them
The most extensive documented case is the supply-chain compromise via ConnectWise in October. Microsoft tracks the actor as Storm-0940, described as aligned with the Chinese state, and says the group used Artex for authentication attempts against more than 2,250 Microsoft 365 tenants, and that it stole customer data. The FBI has, according to Times Now, confirmed Microsoft's findings.
In addition, the advisory points, according to secondary coverage, to several other incidents. India's CERT-In has previously attributed Artex use to the Chinese-linked group "Crafty Falcon," which attacked banks, fintech companies and crypto exchanges and moved money through the victims' own payment systems, the NY Post reports. Barracuda has, according to Times Now, disclosed an Artex-related intrusion via an exposed third-party credential set. South African institutions are not mentioned in the sources and are not relevant here. What both the FBI advisory and WIRED describe is the theft of personal data from thousands of customers of South Korean financial institutions. The number of affected people varies between sources: Times Now states "at least 68,000 people" with reference to the FBI, while Startup Fortune says 67,000. The discrepancy is unresolved; the safest formulation is roughly 68,000, with varying figures reported.
Who the users are is partly unclear. The researchers WIRED spoke to link the usage to hacking activity often carried out by Chinese-speaking cybercriminals, including potentially crypto theft and hunting for exposed login credentials on GitHub. State affiliation is documented in two cases — Microsoft's Storm-0940 and CERT-In's Crafty Falcon — but the split between criminals and state-backed actors is not fully mapped. Some outlets, such as the NY Post, frame Artex as "China-made"; that is broader than what the sources document, which point to specific attributions rather than a general national stamp. (One source, via MSN, claims the developer behind Autumn-27 is the Chinese security engineer Li Puhua, formerly employed at ThreatBook. This is single-sourced and unverified, and is mentioned here only with that clear labeling.)
The economics: nearly free attacks
One of the most striking parts of the coverage is the cost side. Vals's measurements, relayed via Tech.co and repeated by Startup Fortune, estimate that an Artex run powered by DeepSeek costs around $2–3, versus $17–46 on OpenAI's GPT-6-Sol. That is the price of an attempt against a target — a fraction of what manual penetration testing normally requires.
The scale is documented by several companies. Check Point has, according to Startup Fortune, measured more than 400,000 messages against 292 targets in a single campaign, and describes a crypto-focused campaign that consumed 167.5 million tokens without earning anything — a sign that the operators are not all skilled. Amazon says its own decoy-sensor registrations followed Artex agents between October 3 and November 5, catching a campaign that stole API keys to AI services and used them to pay for its own operations. These figures are the companies' own reported findings, relayed through secondary coverage.
Why removing the repository stopped nothing
The developer pulled the original repository from GitHub on October 8, 2026. It helped little. The FBI cites research identifying more than 1,600 exposed Artex consoles still accessible on the public internet, according to coverage of the advisory. Censys counted 1,465 exposed consoles as of November 4, with 32 open directories leaking tools for lateral movement in Windows, a sudo exploit and Linux implants that anyone could download and reuse, according to Startup Fortune. The two numbers need not be in conflict — they likely reflect different dates and methodology — but both point the same way: the tool is still in circulation.
The reason is simple. Artex is open source. Copies, forks, archived versions and a commercial variant, ARTEX PRO, ensure that the framework lives on independently of the original. A GitHub takedown does not remove the tool from the world; it removes only one entry point.
The context: a family pattern
Artex does not stand alone. Google's Threat Intelligence Group named Artex the same week alongside Cairn, HexStrike AI, CyberStrikeAI and PentAGI, describing a movement toward coordinated multi-agent frameworks, according to Startup Fortune, which in turn cites CIO Dive. The pattern the companies describe is consistent: attacks that previously took days are now completed in hours, WIRED writes.
One thing this week's coverage clearly does not say: some outlets have framed this advisory against a series of recent US service outages, but coverage of those outages reports that authorities stated they were not cyberattacks. There is no documentation in the available basis linking Artex activity to those outages, and the two should not be conflated.
What defenders can take from this
The concrete picture drawn by the advisory and the company coverage is not "AI can hack anything," but something more precise: a documented framework, documented incidents and attributed figures. The implications are three. First, Artex lowers the threshold — the FBI warns that hackers with almost no technical skill can run complex operations against many targets simultaneously, which means the volume of attempts can rise even if the attack techniques themselves are not new. Second, the timeline is the new pressure factor: under three hours from targeting to privileged access sets a requirement for automated detection and response where human analysts cannot react manually in time. Third, the Amazon observation shows that the attacks increasingly finance themselves through stolen API keys, which means visibility into one's own AI-service credentials becomes a defensive task on par with passwords and certificates.
The open questions remain many. Who Autumn-27 is, is unresolved. How many were actually affected in South Korea — 67,000 or 68,000 — has not been reconciled. The criminal share of the usage, set against the state-aligned share, is not fully described. And the FBI advisory itself is not publicly available in the material this article is built on; all key figures are attributed through secondary coverage. It is worth remembering as the picture is drawn: this is a threat that is well documented in scale, but still partly unclear in its details.

