For years, the dominant narrative around AI and cyberattacks has been about the front end. Phishing emails that are harder to spot. Social engineering lures that adapt to the target. Better grammar from criminal groups that used to telegraph their origins with broken English.
UAT-10147 is a different story.
What Talos found
Talos recovered approximately 170,000 URLs from an open C2 directory belonging to UAT-10147, giving them an unusually clear view into the operation. The group targets web servers across government, universities, media, technology, and gaming sectors in Brazil, Bolivia, China, Canada, and Vietnam.
The agentic AI integration covers the full post-compromise lifecycle: reconnaissance of compromised environments, payload generation tailored to the target system, exploit automation for secondary access, validation of successful deployment, and persistence mechanisms. Talos recovered AI-generated operational playbooks and troubleshooting logic artifacts from the intrusion chain. The AI is not advising a human operator in real time. The AI outputs are pre-generated and embedded in the attack workflow.
That is what 'agentic' means here, and it matters. An operator who has to consult an AI assistant for each decision is still rate-limited by human throughput. An operator whose tooling incorporates AI-generated playbooks into the attack chain itself can scale the operation without scaling the headcount.
The SPECTRE implant
UAT-10147's primary tool is SPECTRE, a cross-platform implant with Linux rootkit capabilities. It also uses BYOVD (Bring Your Own Vulnerable Driver) techniques for EDR bypass on Windows targets. Supporting tooling includes NoodleRAT, QuasarRAT, Gh0stCringe, and Meterpreter. The combination gives the group coverage across both Windows and Linux server environments, consistent with targeting web servers running mixed infrastructure.
The BYOVD technique involves loading a legitimate but vulnerable driver, then exploiting it to run code at kernel level, bypassing endpoint detection that operates in user space. BYOVD is not new, but its presence in a financially motivated Chinese-speaking group reflects continued democratization of techniques once associated with the most sophisticated actors.
The skill ceiling still applies
Talos published a parallel analysis of AI use across adversary groups, based on artifacts recovered from compromised infrastructure including AI prompt logs. The key finding: most actors can bypass AI guardrails with moderate effort. The amplification is real but asymmetric. A skilled actor using AI moves faster and scales more simultaneous intrusions. An unskilled actor using AI produces more sophisticated-looking failures.
AI tools provide force multiplication at reconnaissance and initial access phases. They do not replace post-compromise operational expertise. UAT-10147 appears to be on the skilled end of the spectrum. They have custom tooling, multi-platform coverage, and a documented target list of 170,000 URLs. The agentic AI integration is not compensating for a lack of skill. It is amplifying existing capability.
What this changes for defenders
The dwell time advantage that detection tools rely on is being compressed. If post-compromise operations run faster because AI handles playbook generation and automation, detection windows shrink. Endpoint telemetry and behavioral detection that assume human-paced lateral movement may need recalibration.
The BYOVD component is a signal to audit which signed but vulnerable drivers are present on critical servers, particularly those not covered by Windows vulnerable driver blocklist updates.
The open C2 directory, while useful to Talos, also means UAT-10147 is not exercising strong operational security. Groups with better OPSEC using similar AI-assisted workflows would leave fewer artifacts. The Talos research is likely seeing one end of a capability spectrum that extends further.
Read the full Talos disclosure at talosintelligence.com. The SPECTRE implant analysis is published separately.
Gigia Tsiklauri is a Security Architect and founder of Infosec.ge. Get in touch if you are tracking AI-enabled threat actors or need help evaluating your detection coverage against BYOVD techniques.