FBI Warns of AI Hacking Tool ARTEX… Even If the Original Comes Down, Copies Keep Attacking
The FBI has issued an official alert over the real-world malicious use of the AI hacking tool ARTEX. On November 13 (local time), the Federal Bureau of Investigation issued a 'FLASH' alert through its Internet Crime Complaint Center (IC3), saying the tool has been used in attacks by various hacking groups since at least August. ARTEX is a tool that, when a target is entered in natural language, uses AI to autonomously carry out everything from system intrusion to data exfiltration; in some cases it took less than three hours from initial intrusion to obtaining administrator privileges.
According to the FBI, ARTEX is a multi-agent platform written in Go that converts natural-language goals into machine commands executed by role-specific specialized agents, minimizing human intervention after the goal is set. It has no model of its own, instead connecting to external large language models (LLMs) such as Claude and GPT, as well as DeepSeek, GLM, Grok, and Kimi. Analysis of attacker-controlled instances showed Chinese-origin models were the most common, with the DeepSeek Flash family particularly prominent. The FBI cited the balance of performance and cost, and the lack of meaningful safeguards against requests framed as penetration testing. The FBI named finance, government, technology, defense, semiconductors, energy, healthcare, manufacturing, telecommunications, and legal sectors as ARTEX targets.
There are cases that illustrate the speed of these attacks. CrowdStrike reported that a China-linked actor it tracks as 'VAPOR PANDA' targeted more than 200 organizations across 14 countries over an 11-day period starting October 17. In one case confirmed by Microsoft, an AI agent went from discovering exposed secrets on GitHub to exfiltrating Azure Key Vault data in just five minutes.
South Korea was also targeted. The FBI said a financially motivated attack against a Korean financial institution exposed the personal data of roughly 67,000 people. There were supply chain attacks as well. Citing a Microsoft report, the FBI wrote that a breach of ConnectWise cloud backups enabled access to more than 2,250 downstream tenants, and Microsoft linked it to the 'Storm-0940' group. However, cybersecurity news outlets cautioned that the figure reflects the scope of targeting rather than confirmed breaches, and that data theft was not proven by successful logins alone.
A developer known as 'Autumn-27' published the ARTEX hacking tool on GitHub in July, then removed the repository on October 8 following reports of attacks on Korean banks. In removing it, the developer wrote, "Considering that the tool has been abused, I will no longer update it and am switching to closed source." However, forks, localized versions, archived copies, and a commercial ARTEX PRO edition continue to circulate, and researchers identified more than 1,600 publicly exposed ARTEX consoles in early November.
The FBI identified edge systems as the vulnerable point. Partner and vendor portals, mobile backends, employee apps, legacy services, and cloud management endpoints were typical, with known weaknesses such as unsigned SAML assertions, XXE, SSRF, and SQL injection chained together into infiltration paths. After intrusion, attackers installed web shells, added credentials to existing OAuth apps, and attempted Kerberoasting, enumerating more than 110,000 Active Directory objects.
In response, the FBI recommended compiling an inventory of internet-exposed applications to reduce the external attack surface, and restricting access to cloud management and network device management endpoints to trusted networks. Countermeasures also include rate-limiting high-impact APIs and finding and rotating secrets embedded in configurations, repositories, and pipelines. Connections to LLM API endpoints and anomalous token consumption are also to be monitored. Suspicious activity can be reported to local FBI field offices or the IC3, with ARTEX configuration files, console screenshots, and service principal login logs cited as useful evidence.
