UAT-10147: cybercriminals turn server attacks into an assembly line with AI

Cisco Talos specialists have released details on a Chinese-speaking cybercriminal group designated UAT-10147. It attacks Windows and Linux web servers worldwide, hitting the education, media, technology and gaming sectors. Most of the victims are located in Brazil, Bolivia, China, Canada and Vietnam. The investigation began after researchers stumbled upon an open directory at 139.180.197[.]150 that was exchanging data with one of the compromised machines.

An industrial approach

The group's defining trait is turning intrusion into a pipeline. The attackers gain initial access through publicly known vulnerabilities, and from there a set of open-source offensive frameworks comes into play: Metasploit, ysoserial, along with the AI tools PentestGPT and DeepAudit. Artificial intelligence is used at various stages — to refine exploits, debug logic, automate post-exploitation, verify that attacks work and even generate accompanying documentation.

The scale is impressive: the open directory held a list of roughly 170,000 URLs, neatly split into 17 files of 10,000 addresses each for convenient processing. The primary targets are the United States, India, the United Kingdom, Germany and the Netherlands. Beyond stealing data, the group engages in SEO fraud, deploying the BadIIS malware on compromised IIS servers — malware distributed under a malware-as-a-service model among several Chinese-speaking groups at once.

SPECTRE and defensive blind spots

The centerpiece of the arsenal is a previously undocumented cross-platform backdoor called SPECTRE, written in C. It was first used in April 2026. It communicates with a command-and-control server over HTTPS and supports up to 45 commands: from keylogging and taking screenshots to shellcode injection and techniques like process hollowing and Early Bird APC injection.

Particularly dangerous is its EDR bypass feature using the BYOVD (bring your own vulnerable driver) method. SPECTRE loads two known vulnerable drivers — MSI's RTCore64.sys and Dell's DBUtil_2_3.sys — and, through targeted writes to kernel memory, unhooks the registered callbacks of security products. As a result, CrowdStrike Falcon, SentinelOne, Microsoft Defender and other solutions stop seeing process and thread creation for the rest of the session.

The Linux version is more modest in its number of commands (29 versus 45), but it carries a trump card — a kernel-level rootkit named Specter, packaged as a loadable module. It gives the attackers persistent control that survives reboots. Talos suspects the rootkit was written by a combination of AI and a human: this is suggested by the detailed comments, decorative function dividers and the many ways of solving a single task — the sort of thing AI tends to generate when asked to “do it thoroughly.”

Blending in with legitimate traffic

A curious technique is exfiltrating stolen data through the legitimate cloud configuration management service Nacos. By polling their own Nacos instance, the attackers check whether victims were successfully compromised without setting up persistent reverse shells or drawing attention with inbound connections. Before connecting, both versions of SPECTRE run sandbox checks: if the weighted score (process name, RAM volume, number of cores, disk space, host names) exceeds 50 points, the malware self-destructs.

Source: The Hacker News