The Chinese-speaking group UAT-10147 has turned artificial intelligence from a code writing assistant into a full-fledged tool for attacks. Attackers instruct AI to look for vulnerabilities, prepare means of exploitation, check their work and help to fix themselves on servers, and the backdoor SPECTRE developed by them is able to disable Windows protection and hide in the Linux kernel.
The group in early 2026 was discovered by Cisco Talos specialists. UAT-10147 attacks the servers of state organizations, universities, media, technology and gaming companies available from the Internet. Confirmed infections were found in Brazil, Bolivia, China, Canada and Vietnam. On one of the servers of intruders, the specialists found a list of about 170 thousand addresses of potential targets.
The main goals of UAT-10147 are to steal data and cheat with SERPs. After entering the system, the attackers install BadIIS or their own handler SeoEngineHandler. Malware replaces the content of sites for search robots and can redirect visitors. One option is specifically focused on Vietnam and recognizes the search engine Cốc Cốc.
For Windows, the grouping uses EfsPotato, GodPotato, JuicyPotato, and RustPotato to increase privileges, adds IIS directories to Microsoft Defender exceptions, creates an administrator account, and secures remote access. QuasarRAT and Gh0stCringe also met in the attacks. On Linux, attackers used known vulnerabilities CVE-2022-0995, CVE-2021-3156, CVE-2015-5287, CVE-2015-3246, CVE-2010-3904 and CVE-2022-0847, after which they installed SPECTRE, Nood
A separate place in the arsenal is written on C backdoor SPECTRE for Windows and Linux. The Windows version supports 45 commands: it can run programs, intercept keystrokes, take screenshots, steal Chrome and Edge credentials, extract registry data, and implement code into processes. Before launch, the malware evaluates the signs of the virtual environment and stops working if the analysis is suspected.
SPECTRE is capable of disabling attack detection tools through vulnerable RTCore64.sys and DBUtil_2_3.sys drivers associated with CVE-2019-16098 (7.8 High) and CVE-2021-21551 (8.8 High). Having been able to change the memory of the Windows kernel, the backdoor disables the functions through which protective products track how processes and streams are created and software modules are loaded.
The SPECTRE version for Linux contains 29 commands and installs a Specter rootkit disguised as the acpi_pad.ko kernel module. Specter hides processes and its own module, increases privileges to root and starts at each system load through the hardware-monitor.service service. To intercept system functions, rootkit uses a full-time ftrace mechanism, which helps to make less noticeable changes to the kernel.
UAT-10147 applies AI and directly during attacks. On the control server of the group specialists found DeepAudit and PentestGPT. The system generated instructions and scripts to exploit ASP.NET View View deseralization errors, checked record rights, diagnosed crashes, downloaded SPECTRE, and created a web shell. The Talos also found signs that the Specter code was being created by the machine, and traces of the AI catalog in an environment where attackers collected their own tools.
According to Cisco Talos, UAT-10147 attacks show that AI is moving from conventional help when writing programs to partially autonomous attacks. The model already helps to explore targets, refine ways of exploitation, look for the causes of mistakes, check the results and prepare tools, reducing the amount of manual work and the level of knowledge necessary for the attacker to carry out a complex attack.
The group in early 2026 was discovered by Cisco Talos specialists. UAT-10147 attacks the servers of state organizations, universities, media, technology and gaming companies available from the Internet. Confirmed infections were found in Brazil, Bolivia, China, Canada and Vietnam. On one of the servers of intruders, the specialists found a list of about 170 thousand addresses of potential targets.
The main goals of UAT-10147 are to steal data and cheat with SERPs. After entering the system, the attackers install BadIIS or their own handler SeoEngineHandler. Malware replaces the content of sites for search robots and can redirect visitors. One option is specifically focused on Vietnam and recognizes the search engine Cốc Cốc.
For Windows, the grouping uses EfsPotato, GodPotato, JuicyPotato, and RustPotato to increase privileges, adds IIS directories to Microsoft Defender exceptions, creates an administrator account, and secures remote access. QuasarRAT and Gh0stCringe also met in the attacks. On Linux, attackers used known vulnerabilities CVE-2022-0995, CVE-2021-3156, CVE-2015-5287, CVE-2015-3246, CVE-2010-3904 and CVE-2022-0847, after which they installed SPECTRE, Nood
A separate place in the arsenal is written on C backdoor SPECTRE for Windows and Linux. The Windows version supports 45 commands: it can run programs, intercept keystrokes, take screenshots, steal Chrome and Edge credentials, extract registry data, and implement code into processes. Before launch, the malware evaluates the signs of the virtual environment and stops working if the analysis is suspected.
SPECTRE is capable of disabling attack detection tools through vulnerable RTCore64.sys and DBUtil_2_3.sys drivers associated with CVE-2019-16098 (7.8 High) and CVE-2021-21551 (8.8 High). Having been able to change the memory of the Windows kernel, the backdoor disables the functions through which protective products track how processes and streams are created and software modules are loaded.
The SPECTRE version for Linux contains 29 commands and installs a Specter rootkit disguised as the acpi_pad.ko kernel module. Specter hides processes and its own module, increases privileges to root and starts at each system load through the hardware-monitor.service service. To intercept system functions, rootkit uses a full-time ftrace mechanism, which helps to make less noticeable changes to the kernel.
UAT-10147 applies AI and directly during attacks. On the control server of the group specialists found DeepAudit and PentestGPT. The system generated instructions and scripts to exploit ASP.NET View View deseralization errors, checked record rights, diagnosed crashes, downloaded SPECTRE, and created a web shell. The Talos also found signs that the Specter code was being created by the machine, and traces of the AI catalog in an environment where attackers collected their own tools.
According to Cisco Talos, UAT-10147 attacks show that AI is moving from conventional help when writing programs to partially autonomous attacks. The model already helps to explore targets, refine ways of exploitation, look for the causes of mistakes, check the results and prepare tools, reducing the amount of manual work and the level of knowledge necessary for the attacker to carry out a complex attack.
