Chinese-Speaking Threat Actor Uses Agentic AI to Exploit Global Web Servers

Cisco Talos identified a Chinese-speaking financially motivated threat actor, tracked as UAT-10147, integrating agentic AI into real-world attacks against internet-facing Windows and Linux web servers.

The actor used AI-driven offensive tooling for vulnerability scanning, exploit selection, exploit validation, payload generation, reconnaissance, persistence and post-compromise operations.

Talos directly observed PentestGPT installed on attacker infrastructure dynamically scanning web servers, executing relevant proof-of-concept exploits, successfully exploiting a website and gathering information from the compromised system.

UAT-10147 also used AI-generated operational playbooks and scripts to validate ASP.NET ViewState deserialization attacks, deploy implants and web shells, conduct reconnaissance and exfiltrate victim information.

Confirmed affected systems were located in Brazil, Bolivia, China, Canada and Vietnam across government, university, media, technology and gaming organizations.

Cisco Talos discovered UAT-10147 targeting vulnerable internet-facing Windows and Linux web servers.

The threat actor maintained a target list containing approximately 170,000 URLs and divided the list into smaller batches for large-scale scanning operations.

Talos observed PentestGPT installed on the actor’s command-and-control infrastructure and used to dynamically scan web servers and execute relevant proof-of-concept exploits.

During observed activity, PentestGPT successfully exploited a website and gathered information about the victim machine using Linux commands.

The actor also used AI-assisted workflows to generate malicious Java deserialization payloads and supporting exploitation tooling.

Recovered AI-generated scripts tested writable directories, validated ASP.NET ViewState remote-code-execution conditions, deployed the SPECTRE implant, installed persistent web shells, performed system reconnaissance and exfiltrated collected information through external web services.

After obtaining access, UAT-10147 deployed malware including SPECTRE, QuasarRAT, Gh0stCringe, Noodle RAT and Meterpreter.

Windows operations included privilege escalation, Microsoft Defender exclusions, scheduled-task persistence and creation of privileged local accounts.

Linux compromises included web-shell deployment, privilege escalation and persistent command-and-control implants.

Talos assesses that the actor’s use of AI progressed beyond simple code generation into iterative exploit refinement, adaptive troubleshooting, exploit validation and post-exploitation automation.

Windows IIS Servers / Linux Web Servers / Web Applications / ASP.NET ViewState / Active Directory / Endpoint Security Controls / Enterprise Credentials

Agentic AI Exploitation / Vulnerability Exploitation / Remote Code Execution / Post-Exploitation Automation

UAT-10147 successfully compromised internet-facing systems belonging to organizations across multiple countries and industries.

Confirmed affected servers were located in Brazil, Bolivia, China, Canada and Vietnam, including systems operated by government, university, media, technology and gaming organizations.

The actor’s infrastructure contained a target list of approximately 170,000 URLs, demonstrating the potential scale of the operation, although this number should not be interpreted as 170,000 confirmed victims.

Successful attacks enabled remote code execution, privilege escalation, persistent web-shell and malware deployment, reconnaissance, credential theft, data exfiltration, defense evasion and long-term command-and-control access.

The campaign demonstrates how agentic AI can reduce the manual effort required to validate exploits and convert successful vulnerability exploitation into repeatable post-compromise workflows.