Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

CloudSEK: Kazakhstan-linked hackers used a DeepSeek-powered AI agent to spy on Central Asian governments

Security firm CloudSEK says a Kazakhstan-linked hacking group used an AI agent running the DeepSeek model to conduct espionage against government targets in Central Asia, according to MLQ.ai. The report places the model inside the attack chain rather than at its margins.

Published

A hacking group linked to Kazakhstan used an AI agent running the DeepSeek model to carry out cyber espionage against government targets in Central Asia, according to research from security firm CloudSEK reported by MLQ.ai.

The notable claim is where the AI sits in the attack chain. The subject is not a person querying a chatbot, but an autonomous agent, with the model acting as the component that executes tasks.

DeepSeek matters here for reasons of availability. Open-weight models can be deployed by an attacker directly, without relying on an external API's access controls and audit logs, which lowers the cost of building a tool that can persist in an environment. The report does not say precisely how the group used the model, whether for phishing text, triage of stolen data or automated reconnaissance.

Attribution deserves caution. The Kazakhstan link is CloudSEK's assessment, and such claims typically rest on infrastructure, code reuse and patterns of behaviour that outside researchers cannot fully verify from a public write-up. It should be read as an allegation awaiting independent corroboration rather than a settled fact.

Even so, the direction of travel is clear: agentic AI is shifting from an enterprise automation tool into a general-purpose capability that both attackers and defenders now use. Detection built around malicious file samples has to extend to model invocation patterns and agent behaviour.

The immediate consequence for Central Asian governments is a higher risk of targeted phishing and internal data loss. For organisations more broadly, the lesson is about cheap replication: once one group proves a path, adoption tends to outrun defensive updates.

What to watch next is whether CloudSEK publishes technical indicators, whether a second research team independently reviews the attribution, and whether open-model providers adjust usage policies or provenance mechanisms around abuse. For now the public record stops at an allegation plus preliminary analysis.

Why it matters

If the assessment holds, open-weight models will keep lowering the technical bar for state-linked cyber operations. Defenders will have to treat the AI agent itself as an observable artefact, not just the malware it produces.

CloudSEKDeepSeekCybersecurityAI Agent
Back to realtime news

Nearby Updates

All