Chinese hackers use DeepSeek AI to boost attacks

Chinese state-affiliated cyber groups have more than doubled their attack volume since incorporating DeepSeek and other open-weight AI models into their operations, according to TeamT5, a Taiwanese threat intelligence firm. The hackers now use AI to handle routine tasks and to develop sophisticated malicious software across multiple stages of an intrusion.

Researchers at TeamT5 told Investing.com that DeepSeek’s offerings are popular among Chinese hackers because of high performance, low cost, and weak built-in safety barriers. The team noted that in many incidents it was not possible to identify which specific AI model produced the code or scripts used in the attack, but recent cases have left clearer fingerprints.

Why DeepSeek appeals to threat actors

The analysts pointed to two practical reasons: low operational cost and relatively weak guardrails. While other Chinese models such as Moonshot’s Kimi K3 are reportedly more capable, they remain too expensive for most cyber operations. TeamT5 said it has not recorded any incidents involving Kimi K3 to date.

Western models remain attractive to attackers, but their safety filters require more effort to bypass, the researchers added. That asymmetry has pushed many Chinese-aligned groups toward DeepSeek, where the cost of producing harmful output is lower.

How the AI is being used in attacks

DeepSeek and similar open-weight models are now deployed across several attack stages, from reconnaissance through exploitation. In recent months, TeamT5 obtained scripts and logs that show Chinese government-affiliated hackers using the model throughout their kill chain.

Specific uses reported by the researchers include:

• Grimfengxi used DeepSeek to create exploit code.
• Huapi used a Chinese AI model, likely DeepSeek, to attack a Taiwanese company’s email system.
• Teleboyi used the platform to collect 1,000 IP addresses from the public internet and map company domains.

What defenders should watch

The findings point to a measurable shift in attacker productivity. Doubling attack volume with the same headcount suggests AI is removing routine bottlenecks, allowing operators to spend more time on the parts of an intrusion that require human judgment, such as lateral movement and credential abuse.

Defenders gain an early edge by hunting for the telltale signs of AI-assisted scripting, including consistent comment styles, characteristic function naming patterns, and artifacts that match known model outputs. Endpoint and network telemetry that flags large-scale automated reconnaissance, such as rapid enumeration of 1,000 or more IP addresses in a short window, can also surface AI-enabled campaigns earlier than traditional indicators of compromise alone.

Performance context for DeepSeek

Separate academic benchmarks of DeepSeek on professional exams illustrate why the model is attractive to technically skilled users. In a comparative study published in the Journal of Medical Systems, DeepSeek-R1 scored 92.0% overall accuracy on 600 multiple-choice questions from the 2024 Chinese National Medical Licensing Examination, compared with 87.2% for ChatGPT-4o, a statistically significant gap. DeepSeek-R1 also outperformed ChatGPT-4o in the low-difficulty subset (95.9% versus 92.0%), with no significant difference in the high-difficulty group.

A second study, published in JMIR Formative Research, compared GPT-4-turbo and DeepSeek-R1 on the 2024 Taiwan Audiologist Qualification Examination, a 300-question test across six subject areas. Both models cleared the 60% passing threshold, with GPT-4-turbo at 80.3% and DeepSeek-R1 at 79.3%. The difference was not statistically significant. Both models handled reverse logic questions well, struggled with complex multiple-choice items, and performed poorly on graph-based questions, with DeepSeek-R1 slightly ahead on graphs at 36% versus 18% for GPT-4-turbo.

Taken together, the benchmarks show DeepSeek matching or exceeding leading Western models on demanding professional tasks, which helps explain why it has also become a capable tool in the hands of state-linked attackers.

FAQ

Which AI model are Chinese hackers using most?

DeepSeek. TeamT5 researchers said DeepSeek is the AI of choice for Chinese hackers because it is relatively powerful, has very low cyber guardrails, and is inexpensive to operate.

How much have attacks increased since AI was adopted?

Chinese state-affiliated attack volume has more than doubled since the groups began incorporating DeepSeek and other open-weight AI models into their operations, according to TeamT5.

What tasks is the AI performing during attacks?

AI is being used across the attack lifecycle, including reconnaissance, exploit development, and attacks on email systems. Documented examples include generating exploit code, targeting a Taiwanese company’s email system, and collecting 1,000 IP addresses to map corporate domains.

---

This article summarizes reporting from yahoo.com.