ARTEX
An open-source AI-driven penetration testing framework that integrates large language models into its decision-making pipeline. It automates the orchestration of attack workflows through a perceive–decide–act loop that dynamically adjusts attack paths, lowering the technical barrier for conducting advanced cyberattacks.
Timeline (last 90 days)
In October and November 2026, Microsoft disclosed cases of different attack groups using ARTEX, including activity tracked as Storm-0940 and Storm-3358; Storm-0940 involved ConnectWise-related supply chain systems, and Storm-3358 was linked to attacks against financial institutions.
On November 13, 2026, the FBI formally issued a security alert about ARTEX; only about four months passed from the public release to a dedicated warning from a federal law enforcement agency.
On October 8, 2026, the original developer stopped publicly providing the ARTEX source code, but ARTEX continued to spread through community forks, archived copies, localized versions and commercialized versions.
The FBI report shows that the models ARTEX supports include Claude, GPT, DeepSeek, GLM, Grok and Kimi, among others; in some tools and attacker deployments, the DeepSeek Flash series is relatively widely used.
ARTEX was first publicly released on GitHub in July 2026 under the developer account name autum27, positioned as an open-source automated penetration testing tool.
In the cryptocurrency attack cases recorded by Check Point, some agents believed they had successfully completed the fund transfer because they interpreted the success message returned by the interface as the actual transfer result, but the researchers did not observe any cryptocurrency arriving in the wallets.
In a representative case, attackers used two ARTEX consoles running continuously for about two weeks, consuming about 36 billion tokens, with model invocation costs estimated at about $700 to $1,500 at then-current API pricing; the agents ultimately obtained more than 300 certificates and associated passwords, but only 3 passed full validation.
Check Point Research found that more than 2,500 systems on the internet had at some point deployed ARTEX consoles, and at the time of the investigation more than 1,600 of those instances were still publicly accessible; among them, 10 instances exposed their complete databases due to misconfiguration.
In some actual intrusions, attackers took less than three hours from selecting a target to gaining privileged access.
ARTEX is a software platform that connects large language models, a task management system and cybersecurity tools; AI analyzes the task, calls tools, reads the results and decides the next step, looping until the task is completed or resources are exhausted.
23 more timeline events
All Facts (20)
The FBI released a 13-page threat bulletin stating that attackers weaponized ARTEX to automatically execute multiple stages of network intrusion, from reconnaissance to data exfiltration, at "machine speed."
60%UnverifiedIn the cryptocurrency attack cases recorded by Check Point, some agents believed they had successfully completed the fund transfer because they interpreted the success message returned by the interface as the actual transfer result, but the researchers did not observe any cryptocurrency arriving in the wallets.
50%UnverifiedCheck Point Research found that more than 2,500 systems on the internet had at some point deployed ARTEX consoles, and at the time of the investigation more than 1,600 of those instances were still publicly accessible; among them, 10 instances exposed their complete databases due to misconfiguration.
50%UnverifiedIn some actual intrusions, attackers took less than three hours from selecting a target to gaining privileged access.
50%UnverifiedThe FBI report shows that the models ARTEX supports include Claude, GPT, DeepSeek, GLM, Grok and Kimi, among others; in some tools and attacker deployments, the DeepSeek Flash series is relatively widely used.
50%UnverifiedARTEX's core program is written in Go, it provides a web management interface based on Next.js, coordinates tasks through multiple sub-agents, and calls external large language models via API; it does not itself provide foundation models and can connect to AI models from different companies.
50%UnverifiedARTEX is a software platform that connects large language models, a task management system and cybersecurity tools; AI analyzes the task, calls tools, reads the results and decides the next step, looping until the task is completed or resources are exhausted.
50%UnverifiedUsing stolen credentials for a cloud backup vendor's application, the attackers launched login attempts against more than 2,250 downstream tenants, with 150,000 failed logins, peaking at 2,900 per minute, and achieved privileged persistence in multiple environments.
50%UnverifiedAn autonomous penetration testing framework called ARTEX (also written ARTX in the video) was released on GitHub by a Chinese author and previously won the Baidu Workshop Challenge.
50%UnverifiedARTEX's architecture uses a graph to organize the penetration process, with one planning agent breaking down tasks and three execution agents invoking real tools to carry them out, with results written back to the asset graph to trigger the next round of planning; the codebase is about 150,000 lines.
50%UnverifiedARTEX does not carry its own large model and instead connects to six external model services, including Claude and DeepSeek.
50%UnverifiedBy analyzing intrusion logs, a security firm concluded that ARTEX has genuine AI capabilities, with evidence including real-time troubleshooting and retries, automatic rerouting in context routing, attack code updated as evidence emerged, and deliberate use of "penetration testing" in comments to bypass model safety guardrails.
50%UnverifiedFrom September to early October, seven South Korean financial institutions were breached, the personal information of about 67,000 people was leaked, and about 200 other enterprises were affected; Claude session histories were left on the attackers' servers, with DeepSeek's Flash series as the primary model used.
50%UnverifiedThe operation codenamed "Zhengqi Panda" covered 14 countries and more than 200 organizations in 11 days; the planning agent for a single intrusion consumed about 1.4 billion input tokens.
50%UnverifiedIn the week ARTEX was taken down, more than 102 forked repositories appeared, and the largest fork had already surpassed the original repository in star count.
50%UnverifiedSomeone implanted a backdoor in a mirrored container, and that backdoored image reached 9,000 downloads.
50%UnverifiedThe number of ARTEX consoles exposed on the public internet worldwide is between 1,400 and more than 1,600; the FBI cited data saying more than 1,600, of which 10 instances had fully open databases, through which operator records were leaked.
50%UnverifiedAgentic frameworks such as ARTEX were designed to improve efficiency, allowing agents to autonomously call APIs, query databases and send instructions.
50%UnverifiedThe attack surface is not the model itself or credential leakage, but the time gap between "the agent decides to act" and "someone discovers what it did."
50%UnverifiedIn the ARTEX attack incident, AI agents operated autonomously without any human review, chaining tool calls across systems and leaving traces only after execution was complete.
50%