Chinese Hackers Use AI Tools to Target 170,000 Web Servers Worldwide

Cybersecurity researchers have identified a Chinese-speaking cybercrime operation tracked as UAT-10147 that's leveraging artificial intelligence tools to attack web servers on a massive scale, with a target list containing roughly 170,000 URLs spanning the globe. Cisco Talos disclosed the threat activity in a two-part report published last week, describing how the group exploits known vulnerabilities in Windows and Linux systems across education, media, technology, and gaming organizations. The attackers have hit victims primarily in Brazil, Bolivia, China, Canada, and Vietnam, though their target list focuses most heavily on the U.S., India, the U.K., Germany, and the Netherlands.

The campaign came to light after researchers found an exposed directory on a compromised server, revealing the attacker's infrastructure and methods. UAT-10147 split its massive target list into 17 smaller files of about 10,000 URLs each to process them more efficiently. The group deploys a mix of open-source hacking frameworks including Metasploit, ysoserial, PentestGPT, and DeepAudit to automate intrusion operations. Once inside victim networks, the attackers install malware for search engine optimization fraud and data theft, often deploying web shells that lead to BadIIS infections and additional backdoors. The operation also uses AI-generated Python scripts for post-exploitation diagnostics, web shell deployment, and blending stolen data with legitimate cloud traffic over HTTPS. On compromised Windows machines, attackers download privilege escalation tools like EfsPotato, configure Microsoft Defender exclusions, delete initial payloads to obstruct forensics, and establish persistence through deceptive scheduled tasks named "Google Chrome Start." Linux intrusions leverage various known local privilege escalation flaws dating back to 2010, followed by deployment of multiple backdoors including Noodle RAT, SPECTRE, and Meterpreter for persistent remote access.

According to Talos researcher Joey Chen, "By routing exfiltrated data to a legitimate cloud-based configuration management service, the attackers effectively blend their traffic with normal administrative operations." The report notes that UAT-10147 employs AI at various attack phases to refine exploits, troubleshoot logic errors, automate post-exploitation workflows, validate exploits, and generate operational documentation. The group has installed PentestGPT, an open-source autonomous pentesting framework, on command-and-control servers to scan web applications and execute proof-of-concept exploits. Researchers also found DeepAudit, an AI-driven vulnerability scanning framework, accessible on the management server, though no evidence surfaced of the tool being used against victims yet—suggesting the attackers may plan to identify vulnerabilities in target environments or audit their own infrastructure against rival threat actors.

The most significant development is SPECTRE, a previously unreported cross-platform backdoor written in C that the threat actor first deployed in April 2026. The Windows version supports 45 commands including keylogging, screenshots, file operations, shellcode injection, process hollowing, and killing endpoint detection and response processes using vulnerable drivers from MSI and Dell. Talos explains that SPECTRE performs targeted kernel writes to unlink registered EDR callbacks, rendering products like CrowdStrike Falcon, SentinelOne, and Microsoft Defender "completely blind" to process creations and other security events. The Linux variant runs anti-sandbox checks and deploys an integrated kernel-level rootkit called Specter that loads as a kernel module, granting persistent kernel-level control that survives reboots and bypasses most user-level security controls. The report suspects the rootkit was developed using a combination of AI-assisted development and human expertise, based on descriptive source code comments, uniform decorative separators explaining each function, and multiple methods to achieve the same purpose—a pattern typical of AI models prompted to be thorough rather than efficient. The operation demonstrates how threat actors are integrating AI to scale intrusion tradecraft and evade modern security defenses, while using legitimate cloud services as asynchronous data collection points to avoid detection risks associated with persistent reverse shells or direct connections. Organizations face a threat that blends commodity malware-as-a-service tools like BadIIS with custom AI-generated utilities and kernel-level implants capable of neutralizing enterprise security products. The erosion of endpoint visibility at the kernel level fundamentally challenges the detection paradigm that most organizations rely on, forcing a broader rethink of defensive architecture beyond signature-based and behavioral monitoring alone.