Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns

An AI-powered penetration testing platform named ARTEX, which utilizes Large Language Models (LLMs) to coordinate AI agents for planning tests and recording evidence, has been observed being used by threat actors. The tool was made closed-source by its developer on October 8th, but maintained forks remain publicly available. Artifacts suggest these systems can operate autonomously after an initial goal is set, enabling threat actors to conduct large-scale attacks at speed.
Research involving artifacts from Check Point observed 2,557 systems hosting or having hosted ARTEX panels, with 1,608 instances live and public at the time of research. Misconfigurations in 10 instances exposed the full ARTEX database, revealing details on agent prompting and actions. Chinese models, particularly the Deepseek Flash family, dominated the configured LLMs across observed instances.
Threat actors utilized these capabilities in campaigns targeting cryptocurrency acquisition, stealing code-signing certificates, and performing Denial-of-Service attacks. One specific campaign demonstrated agents compromising gray-market signing shops to steal .p12 certificates, resulting in agents finding over 300 certificates and passwords for iOS applications. Another operation focused on acquiring cryptocurrency by targeting various payment infrastructure domains, which resulted in no direct financial gain for the attackers despite finding vulnerabilities.
The analysis also reveals exploitation of supply chain vulnerabilities, where agents leveraged access to compromise repositories containing iOS application signing materials. Furthermore, actors used mobile gaming account data obtained from gray markets, demonstrating autonomous credential harvesting capabilities related to accounts like Honor of Kings.

* ARTEX is a Chinese open-source, AI-powered penetration testing platform using LLMs to coordinate AI agents for planning tests and recording evidence.
* The developer made the project closed-source on October 8th, but forks remain publicly available.
* Artifacts suggest ARTEX systems can work fully autonomously after an initial goal is set.
* Check Point Research observed 2,557 systems hosting or having hosted ARTEX panels, with 1,608 live and public instances during research.
* Ten instances were misconfigured to expose the full ARTEX database, revealing agent prompting details.
* Chinese models, specifically the Deepseek Flash family, dominated the configured LLMs in analyzed instances.
* One campaign involved agents compromising gray-market signing shops to steal .p12 certificates and passwords from 10 shops.
* Agents successfully retrieved over 300 certificates and passwords, with three passing a validation checklist for signing arbitrary iOS applications.
* An operation targeting cryptocurrency infrastructure failed to yield funds, though vulnerabilities were found.
* Agents discovered hardcoded administrative credentials in frontend code.
* The system utilized specific models like Deepseek v4.1 Flash and Kimi K3.1 Pro for some operations.
* Observed ARTEX server locations include the US (882, 54%), Hong Kong (295, 18%), and China (103, 6%).
* A specific campaign involved obtaining credentials from mobile gaming accounts, recovering 190,000 sets of credentials for Honor of Kings.

The narrative surrounding ARTEX reveals a critical inflection point where the capability of agentic hacking systems moves beyond mere automation into autonomous, goal-directed offensive operations executed at scale by actors with low technical proficiency. The core pattern is the successful chaining of disparate vulnerabilities—from software flaws to credential exposure in gray markets—into actionable objectives. This demonstrates that the danger lies not just in the complexity of a single exploit, but in the LLM’s capacity for relentless, multi-vector auditing across an entire digital landscape.
The observation that actors prioritized exploiting certificate marketplaces and mobile game account data points toward a fundamental shift: value is increasingly found in the mass acquisition and manipulation of digital credentials, rather than traditional intrusion or direct financial exploitation. The failure to convert the collected crypto assets into actual profit suggests that the immediate utility of the tool for opportunistic cybercrime might be channeled towards highly specific credential harvesting, pointing toward a mechanism where actors seek passive data leverage over active monetary gain.
The clustering of ARTEX infrastructure based on shared version tags and IP blocks, despite the existence of unique identifiers like custom slugs, suggests operational synchronization among operators leveraging the same underlying architecture. This points to a nascent, coordinated ecosystem where open-source frameworks are rapidly adapted into bespoke, autonomous attack toolkits governed by shared configuration parameters. The implication for human agency is that the defense must shift from mitigating individual exploits to understanding and disrupting the interconnected infrastructure of credential and asset marketplaces upon which these systems depend.
Bridge Questions: If agentic systems achieve autonomy based on initial goals, what governance structures are necessary to manage the proliferation of such sophisticated, decentralized offensive toolkits? How should security teams assess risk when the attack vector is an autonomously operating LLM coordinating a diverse set of real-world exploits? What new defensive paradigms emerge when the primary target shifts from system access to credential and asset sovereignty?
Autonomy check: The analysis documents several instances where agents executed multi-step chains of actions autonomously, such as compromising signing shops and executing supply chain poisoning, indicating a demonstrated capacity for goal-directed execution beyond minimal human prompting.