SecurityNews

State-linked hackers turn to Chinese AI models

Hacking groups linked to China, Russia, North Korea and Iran are using AI for reconnaissance, malware development and social engineering. According to research by TrendAI, Chinese actors switch to domestic models as soon as access to Western models is lost. Russian malware also makes use of a Chinese model.

This is revealed in the report From State AI Ecosystems to Global APT Cyberthreats by TrendAI. The researchers mapped how APT groups from the four countries use AI, from adapting exploits at scale to deploying deepfake identities in employment fraud. Western AI providers are trying to restrict these groups’ access to their models. According to TrendAI, China has now built its own AI ecosystem, which China-linked actors are already actively using.

Chinese groups switch models mid-attack

TrendAI says Chinese cyber actors are now structurally integrating domestic AI into the attack process, whereas previously it went no further than experimentation. In a recent attack that the researchers attribute to the APT group Earth Lamia, the attacker switched during an ongoing intrusion from the Western model Claude to the Chinese MiMo-V2.5-Pro when access to Claude became unreliable. According to the report, the groups Earth Krahang and Earth Naga use AI to adapt exploits at scale and to iteratively generate and rewrite malware code.

Russia, North Korea and Iran

According to TrendAI, a country does not need to be a frontrunner in AI development to exploit AI in cyberattacks. Russia compensates for limitations in its own AI infrastructure by reusing existing models. The researchers point to the malware LAMEHUG, which uses a public API to the Chinese model Qwen to generate commands on a compromised system.

North Korea uses AI to obtain money and trade secrets through IT workers. Operatives use deepfakes, AI-generated CVs and live support from language models during job interviews to gain entry to Western companies, including AI companies. According to the report, Iranian groups such as APT42 rely on commercial models such as Gemini for personalised phishing emails, debugging malware and targeted disinformation campaigns.

Feike Hacquebord, Principal Threat Researcher at TrendAI and co-author of the research, says state-linked attackers do not need the latest models to benefit from them. In his view, it is mainly about access to the technology. Hacquebord indicates that AI is already being used to accelerate reconnaissance, social engineering, malware development and post-compromise activity. Defenders, he argues, should focus less on the model an attacker uses and more on the speed, scale and automation that AI adds to existing attack techniques.

Implications for detection and attribution

According to the report, fully autonomous AI attacks are not yet occurring on a large scale. Humans still determine targets and strategy, although the line between human direction and execution by AI is blurring. TrendAI sees three shifts for defenders.

Malware that connects live to external AI APIs can be identified through network traffic. When attackers use locally hosted models, as North Korea does with Ollama according to the researchers, that visibility disappears. Detection then shifts to behavioural analysis on the endpoint.

In addition, attribution is becoming more difficult. Because code and text are increasingly written with generative AI, hackers’ stylistic traits are losing their value. TrendAI advises paying closer attention to infrastructure and modus operandi. Finally, the researchers note that Russia, China, North Korea and Iran are cooperating more on AI, meaning knowledge, R&D and tools are being exchanged between these countries more quickly.

TrendAI’s advice is to put less effort into identifying specific models and more into endpoint detection and keeping pace with the speed of AI-driven attacks.

The full report can be read at From State AI Ecosystems to Global APT Cyberthreats.

This article was originally published in Dutch on MSP Business. Read the original →