Investigators say experienced operators rely on certain China AI models with limited safeguards to automate reconnaissance and craft sophisticated malicious software

According to new research from cybersecurity firm TeamT5, Chinese state-affiliated hacking groups have more than doubled the volume of their cyberattacks since integrating DeepSeek and other open-source AI models into their operations.

The findings, reported by Yahoo and other outlets, show that experienced threat actors are now using relatively inexpensive AI tools with limited safety restrictions to handle routine reconnaissance work and even develop more sophisticated malicious software.

TeamT5 noted that while Western AI models remain highly sought after, their stricter safety measures require considerably more effort to bypass, making them less attractive for large-scale malicious campaigns. Although other Chinese models such as Kimi K3 offer greater capabilities, they remain too costly for hackers to deploy broadly, leaving DeepSeek as the preferred option for its combination of performance, affordability, and customizability.

The research identified several specific groups leveraging the technology.

• A group called Grimfengxi had used DeepSeek to generate exploit code
• Another group known as Huapi had employed a Chinese AI model believed to be DeepSeek in an attack targeting a Taiwanese company’s email system. 
• A third group, Teleboyi, had used the platform to gather approximately 1,000 IP addresses and map corporate domains as part of its reconnaissance efforts. 
• Threat group Slime22 had used Claude Code after breaching a Taiwanese technology firm, posing as cybersecurity engineers to circumvent the model’s safeguards.
• A threat actor tracked under the aliases knaithe and KnYuan had used DeepSeek through the open-source Hermes Agent framework to target more than 460 systems, successfully breaching at least three organisations by exploiting a Citrix NetScaler vulnerability. The operator had also attempted to use Claude Code and OpenAI’s models, but safety measures had blocked the malicious requests and ultimately led to the account being disabled. DeepSeek, accessed through an open-source framework without client-side restrictions, had proceeded without hindrance, autonomously conducting vulnerability enumeration, downloading public exploit code, and launching attacks with minimal human intervention.

Researchers have cautioned that, while not every intrusion could be tied to a specific AI system, DeepSeek’s offerings have become especially popular among hackers in the region.

Not just Chinese AI models involved

Other researchers warn that large language models can accelerate vulnerability research, target prioritization, and exploitation workflows, underscoring the growing threat posed by AI-assisted cyberattacks.