FBI Warns: Free AI Hacking Tool Artex Spirals Out of Control, Breaches Cloud Systems in Three Hours

The free open-source AI penetration tool Artex is being weaponized by criminals and state-backed hackers, and what gets breached are often forgotten flanking systems.
A joint FBI and Microsoft investigation reveals that the open-source AI penetration testing tool Artex is being widely abused by criminal groups and state-linked hacking organizations. The tool connects to large language models, allowing users to describe targets in natural language and automatically advance numerous testing steps. Real cases include the leak of information on approximately 67,000 people at South Korean financial firms, and a state-linked organization generating over 150,000 failed logins in five days, peaking at 2,900 attempts per minute. However, Microsoft observed no successful logins; the actual breakthrough came from leaked credentials in the ConnectWise backup application, with attackers leveraging a trusted app to reach over 2,250 customers. The article notes that most breached systems are forgotten flanking systems—partner portals, app backends, old services—rather than main sites; meanwhile, over 1,600 Artex control panels are exposed to the public internet, showing attackers are equally careless. The fix is not novel: inventory exposed surfaces, enable second-factor authentication, and review third-party app permissions.
A free, open-source AI tool originally designed to help businesses audit their own vulnerabilities is being weaponized by criminal groups and state-backed hacking organizations. A joint investigation by the FBI and Microsoft reveals a troubling fact: attack speeds have reached machine level, but what gets breached are often forgotten "side doors."
What Artex Is: From Penetration Testing Assistant to Attack Engine
Penetration testing is a profession that has existed for decades: companies pay security experts to simulate intrusions into their own systems, finding vulnerabilities before the real thieves arrive. This work is slow and meticulous, and qualified practitioners are scarce and expensive.
Artex changes the cost structure of this process. Its developers released it as free open-source software that anyone can download and run. It doesn't include a built-in AI model itself; instead, it connects to existing large language models—like an engine that can be fitted into different vehicle models. Users simply describe their target in natural language, and a group of AI assistants automatically advances a large number of testing steps around the clock.
According to the FBI, Artex accumulated approximately 1,500 to 2,000 GitHub stars in ten weeks. The developers stated its purpose is for learning and research, helping organizations test systems they have authorization to test. But a tool that can find your own open door can equally find someone else's.
The Boundary Between Penetration Testing and Red Team Tools
Penetration testing tools themselves are not illegal. Commercial and open-source frameworks like Metasploit and Burp Suite have existed for years, and security professionals use them to verify defense effectiveness. What makes Artex special is two things: first, it delegates "orchestration and decision-making" to large language models, lowering the barrier for users to understand vulnerability principles and exploitation chains; second, it is completely open-source and freely distributed, making it impossible to restrict usage through licensing or account systems. The industry typically describes such software as "dual-use tools"—with the same code, the difference between authorized testing and unauthorized intrusion lies solely in whether written permission has been obtained from the target. The FBI's warning is not directed at the tool itself, but at the reality of its large-scale abuse, which also highlights the structural problem that open-source security tools cannot be recalled once the publisher takes them down.
Real Attack Cases: Loud, But Too Fast to React
According to a report from security firm CrowdStrike, someone used Artex to breach South Korean financial companies and steal data, affecting the personal information of approximately 67,000 people. The developer stated the next day that they were not involved and took down the project, but free software cannot be recalled once it spreads—copies, modified versions, archived downloads, and even a paid "Artex Pro" version continue to circulate.

A larger-scale case comes from Microsoft's tracking. A hacking group described as linked to the Chinese government has been using Artex against hundreds of organizations worldwide since September. In one October campaign, the tool launched credential-stuffing attacks against hundreds of work accounts at eight organizations, generating over 150,000 failed logins in five days, peaking at 2,900 failed attempts per minute.
Interestingly: Microsoft did not observe a single successful login; the vast majority of activity merely triggered account lockout mechanisms. Machine speed cannot directly prevail against a lock that "knows when to say no."
But the same organization found another path during the same period. Many enterprises use ConnectWise Cloud Backup to back up work emails and files, and this application is granted permission to access customer accounts—equivalent to giving the cleaning company the keys to your home. In October, ConnectWise discovered that the application's own credentials (its master key) had been stolen. Attackers leveraged this application that customers already trusted to attempt logins to over 2,250 customers' cloud accounts. Most attempts failed, but Microsoft stated that the organization still maintains deep, persistent access in many organizations, sometimes taking less than three hours from initial probing to success.
This is not an isolated case. Microsoft also reported that a Cambodian criminal group used Artex to attack banks and financial firms in Southeast Asia; cloud security company Datadog observed it using stolen access keys to search for financial data in multiple enterprises' Amazon cloud accounts.

What Is "Credential Stuffing" and "Account Lockout"
The "stuffing" mentioned in the article refers to credential stuffing: attackers use username/password combinations leaked from other sources to attempt bulk logins to target systems, betting that users reuse the same password across multiple platforms. The scale of 2,900 attempts per minute is precisely the core advantage of automated tools over manual attacks. Account lockout is a common defensive strategy: after several consecutive incorrect passwords, the account is temporarily frozen, making brute-force guessing infeasible in terms of time cost. This also explains why Microsoft observed 150,000 failed logins with zero successes—the lockout strategy offsets the advantage of machine speed. But locking itself has costs: attackers can deliberately lock out large numbers of real users, creating denial-of-service-style disruption. Therefore, modern defense emphasizes passkeys and other phishing-resistant authentication rather than relying solely on passwords plus lockout.
Attackers Also Forgot to Close Their Own Doors
The most unexpected part of the story is that the people using this AI hacking tool also left their own doors wide open.
Artex runs through a web control panel where operators watch the AI assistants work. This panel should never have been exposed to the public internet. But according to the FBI, researchers discovered over 1,600 control panels exposed online in early November; scanning firm Senses found that hundreds of them didn't even have passwords set. Of course, some panels belong to legitimate security testers.
Carelessness is bidirectional. Part of the reason CrowdStrike could attribute the South Korean attack to a specific hacker was that this person had publicly exposed a batch of their own files on the internet.
What Actually Gets Breached Is the Door You Forgot
The FBI's list is very specific: the risk is not in the main site that everyone watches, but in flanking systems—partner and supplier portals, hidden backends of mobile apps, old services nobody shut down, tools built by some department and forgotten by everyone. The FBI listed ten targeted industries, including banking, government, technology, defense, healthcare, telecommunications, law firms, and energy and manufacturing. But to be clear, the alert did not describe attacks on power grids or factory machinery. Security teams at EU institutions also issued independent warnings.
What this means for ordinary people: those 67,000 South Korean victims never touched Artex; their information simply happened to be stored at the breached company. When a supplier like a backup service is hit, every customer who trusts it instantly enters the blast radius.
The Risk of Third-Party App Authorization and OAuth Tokens
The danger of the ConnectWise Cloud Backup breach lies in the design of authorization mechanisms like OAuth: after an enterprise administrator once consents to a third-party application accessing emails and files, that application holds a long-lived token and can act on their behalf without the user re-entering a password. Once attackers steal this token or the application's own service credentials, they bypass all employee-facing authentication measures of the victim enterprise—because the request appears to come from a trusted application. Such attacks have been called "trust chain attacks" or "token theft" in recent years. Defense priorities include: regularly auditing authorized third-party applications, adopting conditional access policies to limit the scope of data applications can reach, shortening token validity periods, and supporting rapid revocation. For individual users, the "connected apps" list in Google or Microsoft accounts is the household version of the same attack surface.
Why It's Not a Super Hacker, But a Tireless "Doorknob Testing Machine"
After analyzing Artex-related attacks, TechCrunch pointed out that many intrusions rely on simple weaknesses: guessing weak or known passwords, accounts lacking second-factor verification. The Datadog case began with leaked keys, and the ConnectWise attack rode on a trusted application.
The FBI also noted that the tool is "noisy"—generating enough activity to trigger alarms; the problem is that it leaves defenders too little time to react.

A detail worth noting: in the attackers' copies, low-cost AI models do the main work, chosen partly because these models do not display visible safety restrictions when tasks are described as "security testing."
Differences in Safety Guardrails Across Low-Cost Models
Mainstream commercial large model vendors typically add refusal policies at the model level: when a request involves unauthorized intrusion, vulnerability exploitation, or malicious code, the model will refuse to cooperate. However, some open-weight models or low-cost models obtained through specific fine-tuning or jailbreaking have weaker safety alignment, and are more likely to output actionable attack steps, especially when tasks are packaged as "security testing" or "educational research." This explains the detail that "one reason for choosing low-cost models is that they do not display visible safety restrictions." The implication for defenders is: as weakly aligned models proliferate, the barrier to attack automation will continue to fall. Security assessments should not assume adversaries are constrained by model vendors' refusal policies, but should assume attackers can always find a model willing to cooperate.
What to Do: Enterprise Checklist and Personal Habits
For enterprises, the FBI's recommendations are quite clear: inventory all internet-facing systems, including forgotten side doors, and shut down unnecessary ones; move control panels and admin pages off the public internet; prioritize patching internet-facing systems; enable second-factor verification for all important accounts; review which external applications and suppliers can reach your accounts and remove those no longer in use; search for passwords and keys scattered in code and files and rotate them; monitor for sudden spikes in failed logins; and report anomalies through ic3.gov (US) or your national cybersecurity agency.
For individuals, the FBI alert contains no recommendations, but the following practices map one-to-one to these attack paths: use unique passwords for each account; enable passkeys or authenticator codes for email, banking, and social accounts; check connected third-party apps in Google or Microsoft accounts and remove those no longer in use—this is the household version of the ConnectWise lesson; keep automatic updates enabled for phones, computers, and home routers; when notified that your data is involved in a breach, immediately change passwords, log out of all devices, and remain vigilant about emails and text messages that "seem to know your information."

If you run a small business, directly ask your IT and backup service providers whether they are affected, and clean up departed employees' and old suppliers' accounts today.
Conclusion: The Fix Is Equally Boring
Artex did not invent new intrusion methods. It turned the slow work of checking every door one by one into a fast, cheap, and tireless automated process. The corresponding fix is also boring: fewer forgotten doors, better locks, and one extra check on the important doors.
The question worth keeping is: in your own digital life, which door have you forgotten?
Related articles

Anthropic's IPO Eve: Five Risks Behind the Trillion-Dollar Valuation and the Truth About Its Governance
Breaking down Anthropic's IPO prospectus: $184M daily revenue at a trillion-dollar valuation, the truth behind $104B in paper losses, a five-class share governance structure, unusual disclosures in 94 pages of risk factors, and how the lock-up schedule may affect retail investors.

FBI Warns: Hackers Weaponize AI Agent ARTEX to Attack 200+ Organizations
FBI warns that AI penetration framework ARTEX is being used to attack 200+ organizations, with CrowdStrike tracking the VaporPanda group. We break down ARTEX's agent orchestration architecture, the real attack chain, loss data, and the FBI's five defense recommendations.

Verse Engine: Turn Blender/Unity/Unreal Scenes into Browser 3D Links in One Click
Verse Engine by Ctruh is a browser-native 3D and XR engine that publishes Blender, Unity, and Unreal scenes as browser links in one click. View-based streaming opens a 1.03 billion-triangle city scene on a laptop. It offers embeds, API, and SDK for developers, 3D artists, architecture, and e-commerce teams.