AI Hacking Tool ARTEX Survives GitHub Removal; FBI: Some Intrusions Reach Privileged Access in Hours | #hacking | #cybersecurity | #infosec | #comptia | #pentest | #hacker


The FBI warned on Friday, Nov. 13, that criminals and state-aligned hackers are running ARTEX, a free AI penetration-testing framework, and that some of their intrusions reached privileged access in under three hours. The bureau also says the withdrawal did not stop spread of the tool, so any organization with partner portals, mobile backends or forgotten cloud apps now faces automated attacks on exactly the systems it watches least.The alert also points to a control that copying code cannot remove: every ARTEX agent must call an outside large language model, and the FBI lists several providers’ API domains as indicators worth investigating.FBI Says Withdrawing ARTEX From GitHub Did Not Stop Forks, Archived Copies or ARTEX PROA developer using the handle Autumn-27 released ARTEX on GitHub in July as a tool for authorized security testing. On Oct. 8 the developer posted a statement about abuse and said the project would become closed source, with no further public updates. The FBI says malicious use goes back to at least August, and that community forks, localized variants, archived copies and a commercial edition called ARTEX PRO kept the tool available.That outcome follows from how open-source licensing works: free and open-source software can be legally forked without prior approval, so a maintainer cannot recall code that has already been distributed. One fork, AntiFox, began as a copy of ARTEX version 0.1.1 and remains available on GitHub as version 2.0. Arctic Wolf researchers tied AntiFox interfaces hosted on Alibaba Cloud to attacks that included AWS secret harvesting, according to a Nov. 11 BleepingComputer report on AntiFox.Copies are also a hazard to the people who run them. A researcher using the alias Badbyte reported a copied ARTEX repository pointed to a booby-trapped Docker image that gave anyone with a secret trigger a root shell, and that 62 internet-facing consoles served the backdoored build between Nov. 2 and 4. Public reporting does not say whether the backdoor was ever used.How Does ARTEX Reach Privileged Access in Under Three Hours?ARTEX is a multi-agent system written in Go with no model of its own. A user states a goal in plain language, and a team of AI agents plans and runs the attack steps, recording goals, facts and findings in a shared graph. Microsoft says multiple ARTEX-driven attacks gave the actor privileged cloud access in less than three hours from initial targeting.The techniques are not new. The FBI alert lists SQL injection, in which crafted input alters a database query, and server-side request forgery, in which an attacker tricks a server into fetching internal resources such as cloud credentials. It also lists authentication bypasses, hardcoded secrets in JavaScript bundles, web shells, and Active Directory attacks known as Kerberoasting and DCSync. What changes is tempo. The FBI describes the tool as noisy, with high-volume traffic that can trigger alerts, but says its speed leaves defenders little time to respond.CJ Moses, Amazon’s chief information security officer, reached a similar conclusion on the same day. Amazon’s decoy sensors, set up to look like model-hosting services, captured ARTEX agents’ own operating memory between Oct. 3 and Nov. 5, and Amazon said none of the techniques in the campaigns were new. Matt Snyder of Aviatrix told Dark Reading that operational security around many deployments is weak even though the framework is sophisticated, which fits how many deployments sit exposed on the internet.FBI and Microsoft Describe 2,250 ConnectWise Tenants DifferentlyThe FBI alert says one supply-chain compromise enabled access to more than 2,250 downstream tenants of ConnectWise Cloud Backup. Microsoft’s own report is more careful. It says the actor, Storm-0940, used ARTEX to attempt sign-ins to more than 2,250 tenants, that most of the targeting failed, and that Storm-0940 still gained persistent privileged cloud access at many organizations. Microsoft adds that an affected application or successful authentication does not prove data theft.The distinction matters to ConnectWise customers deciding whether to open a breach investigation. Storm-0940 is a China-state-aligned actor active since 2021, according to Microsoft, and its peak rate reached about 2,900 failed sign-ins per minute during a password-spray campaign. The attack worked around multi-factor authentication because the stolen credential belonged to an application, not a person, so there was no interactive sign-in to challenge. OAuth, the open standard for access delegation, lets apps act without a user’s password. ConnectWise suspended several services, and its advisory as of Nov. 4 says Cloud Backup is partially restored in phases, with partners required to re-authorize each Microsoft 365 tenant to a replacement app.Other campaigns show the same tool in unrelated hands. CrowdStrike reported on Oct. 7 that an unattributed, likely Chinese-speaking and financially motivated actor used ARTEX against South Korean financial firms, and the FBI puts that theft at about 67,000 people. Microsoft separately tracks Storm-3358, a Cambodia-based, financially motivated cluster that began using ARTEX against financial organizations in October, and the FBI lists its infrastructure as active on Nov. 6. No vendor has published evidence linking these actors to one another beyond shared use of ARTEX.Censys, Check Point and FBI Count Exposed ARTEX Consoles DifferentlyInternet-scanning firm Censys counted 1,465 hosts running an exposed ARTEX console on Nov. 4, with 726 in the United States, 387 in Hong Kong and 177 in China. About 650 of those sit in two identically configured hosting blocks, which leaves roughly 815 apparently distinct deployments, and 155 hosts also ran other attack tooling such as Cobalt Strike. Check Point Research found 2,557 systems that hosted ARTEX panels, of which 1,608 were live. The FBI puts public exposure at more than 1,600 consoles as of early November.None of these figures isolates criminal use. Censys says many instances are likely used by authorized testers, and the console does not need to face the internet at all, so total use is likely higher than any scan shows. The FBI alert does not explain why its figure and Censys’s differ; differing scan methods and dates probably account for the gap.Forks Escaped Developer, but Every ARTEX Agent Still Needs LLM API AccessAccording to The Record’s account of the alert, ARTEX can integrate Anthropic’s Claude, OpenAI’s GPT, DeepSeek, GLM, Grok and Kimi. The alert lists outbound calls to API domains for Anthropic, DeepSeek, GLM and xAI as indicators to investigate alongside other suspicious activity, not to block on sight. It also tells defenders to look for self-hosted gateway and reseller software, which operators can use to route model access.Amazon found that gateway and AI-application servers are themselves targets: in one campaign, ARTEX agents hunted for exposed model-provider keys, validated them and pooled them, in what Amazon called a self-sustaining operation in which stolen keys paid for further credential hunting. Amazon assessed with moderate confidence that disparate Chinese-nexus groups were behind these campaigns.A second Amazon-observed campaign tried to defeat identity checks at two cryptocurrency exchanges with real-time deepfake video and forged IDs. In the recorded attempts against one identity vendor’s liveness check, the vendor’s server rejected the deepfake every time, and Amazon says it did not observe the activity in real time and cannot verify what the agents reported. Amazon advises organizations to watch for AI service misuse, such as unexpected spend, so model-access logs and AI spending now belong in security monitoring because they show an agent at work even after code has been copied beyond any developer’s reach.Chinese Models Dominate Attacker ARTEX Instances; FBI Cites Weak Observable GuardrailsThe FBI says attacker-controlled ARTEX instances mostly use Chinese models, and that DeepSeek’s Flash family is the most prevalent because of its capability balance, low cost and lack of observable guardrails when requests are framed as penetration testing. Check Point’s analysis, as summarized by Let’s Data Science, found American models on only three instances and said guardrails triggered only on Anthropic’s Opus 5.5, after which the actor moved to more permissive Chinese models. Coverage of the report notes that these observations, drawn from 10 analyzed instances, do not establish a general comparison of model safety. Western tools are not irrelevant to attackers: CrowdStrike found Claude Code session histories in the South Korea actor’s exposed directories, alongside DeepSeek as the primary ARTEX backend.No independent testing of the Flash family’s guardrails has surfaced in public reporting, so the FBI’s characterization stands as the bureau’s assessment. The closest evidence is a 2025 evaluation by the Commerce Department’s Center for AI Standards and Innovation, which tested older DeepSeek models. It found they trailed leading U.S. models on almost every benchmark and were far more susceptible to agent hijacking, and TechRepublic’s report on that evaluation said DeepSeek complied with public jailbreak prompts in 95 to 100 percent of tests, against 5 to 12 percent for U.S. models.Chinese Model Providers Operate Under National Intelligence Law Cooperation DutiesAnyone weighing a Chinese model API for their own work, whether for security testing or anything else, faces legal duties that no product setting changes. DeepSeek is headquartered in Hangzhou and owned by hedge fund High-Flyer. China’s National Intelligence Law of 2017 says in Article 7 that organizations and citizens must support and cooperate with national intelligence work “in accordance with law.” Scholars dispute how far that duty reaches: one reading calls it defensive, another says it shifts obligations toward offense, and a 2024 analysis says it is unclear whether Article 7 was meant to require active information gathering, while noting that none of that limits the state’s espionage capabilities. Whether it reaches overseas subsidiaries is disputed between law firms: a 2019 Mannheimer Swartling report said it could, while a Zhong Lun opinion for Huawei said it does not.Two other laws add to it. The Data Security Law of 2021, in force since Sept. 1, 2021, requires Chinese government approval before organizations provide mainland-stored data to foreign courts or law enforcement, and imposes localization rules on its highest tier of “core data.” The Cybersecurity Law of 2017, effective June 1, 2017 and amended in 2025, requires network operators in Article 28 to provide technical support and assistance to public security organs. Those duties are set by statute, not by a company’s published privacy policy. No independent audit of what any Chinese model provider has handed to authorities has surfaced in public reporting. Self-hosting open weights and keeping credentials and internal data away from any model harness reduces exposure, but no technical measure removes the underlying legal obligation.Defenders Should Inventory Partner Portals and Treat LLM API Calls as SignalsThe FBI’s recommendations start with attack surface. Inventory every internet-facing application, including partner, supplier and business-unit systems. Keep cloud administration, Kudu and SCM, hypervisor and backup interfaces off the public internet, and put enterprise apps behind zero-trust access where possible. Apply least privilege to multi-tenant app permissions and consent grants, and watch for credential additions, role assignments and application-only activity.Organizations that installed the affected ConnectWise apps should review tenants by application ID, not display name, and look back to at least Oct. 1. The bureau says not to delete enterprise apps as a first response, because that destroys evidence, and to keep affected ConnectWise connections disabled until the vendor issues restoration guidance; ConnectWise has since begun issuing that guidance.On portals and APIs, rate-limit high-impact functions, alert on sequential lookups and unusually high record counts per session, enforce authorization on the server, and never show stack traces or SQL errors to unauthenticated users. Hunt for secrets in repositories, container images, build artifacts and pipelines, and rotate any that turn up. The FBI cautions that cloud IP addresses get reassigned, so check indicators against current telemetry before blocking. Victims should report to a local FBI field office or the Internet Crime Complaint Center at ic3.gov.The decision framework for readers comes down to four checks. ARTEX is free and no longer recallable, so assume attackers have it. Its speed makes the forgotten partner portal the likeliest entry point. Model-access logs and unexpected AI spending are early signals that survive forks. And any plan to route data through a Chinese model API should account for thinner independent evidence on its security, weaker observed guardrails and the cooperation duties its provider carries under Chinese law, not only its price.Frequently Asked QuestionsWhat is ARTEX AI?ARTEX is an open-source, Go-based penetration-testing framework in which AI agents plan and run intrusion steps, from reconnaissance to data theft, with little human input. It connects to outside language models instead of shipping its own, according to the FBI.Is ARTEX still available after the developer pulled it from GitHub?Yes. The FBI says community forks and ARTEX PRO, along with localized variants and archived copies, kept it in circulation. Censys counted 1,465 exposed consoles on Nov. 4.Can blocking AI model API domains stop ARTEX attacks?No. The FBI lists calls to API domains for Claude, DeepSeek, GLM and Grok as indicators to investigate with other suspicious activity, not as grounds to block on sight. Operators can also route requests through gateways and resellers, so treat the calls as one signal among several.Does the FBI alert mean ConnectWise customers were breached?Not necessarily. The FBI describes access to more than 2,250 downstream tenants, while Microsoft says the actor attempted sign-ins to more than 2,250 tenants and that most targeting failed. Customers should check Entra logs by application ID rather than assume either outcome.

Click Here For The Original Source.

——————————————————–

............