Chinese State Hackers Doubled Their Attack Volume After Wiring DeepSeek Into Their Operations

Chinese state-affiliated hacking groups have more than doubled their attack volume since integrating AI models into their workflows, according to new research from TeamT5, a Taiwanese threat-intelligence firm cited by Bloomberg. The model of choice: DeepSeek.

The reasoning is blunt. DeepSeek is powerful, free to run locally, and — in TeamT5’s words — comes with “very low cyber guardrails.” Open weights mean no usage monitoring, no account bans, no safety team reviewing abuse reports. For a state-backed operator, that combination beats any frontier API.

The report names names. A group tracked as Grimfengxi used DeepSeek to generate exploit code. Teleboyi used it to harvest 1,000 IP addresses and map a target’s domain infrastructure — the tedious reconnaissance work that used to consume analyst hours. Huapi hit a Taiwanese company’s email system with malware researchers believe was built with a Chinese model, most likely DeepSeek.

The most striking case involves a Western tool. TeamT5 says a group tracked as Slime22 breached a Taiwanese technology firm, installed Kali Linux inside the compromised environment, and then directed Claude Code — Anthropic’s coding agent — to run lateral movement across the network. Attackers are not loyal to any one vendor. They use whatever works, Chinese or American, and they mix tools mid-operation.

Why the doubling matters

The volume increase does not come from smarter attacks. It comes from delegation. AI now handles the mundane middle of the kill chain — writing variants of malicious scripts, parsing scan output, drafting phishing lures — so a fixed number of human operators can run twice as many concurrent campaigns. That is the same productivity story every legitimate engineering team tells about AI coding tools, pointed in the opposite direction.

Taiwan sits at the sharp end of this. The island absorbs millions of probes daily, and TeamT5’s visibility into Chinese operations is unusually deep because so many of the targets are Taiwanese firms. But the report’s examples span targets abroad, and the tooling travels: an exploit generated against a Taiwanese email server works just as well against the same software running anywhere else.

What to do about it

There is no CVE to patch here — the threat is capacity, not a single flaw. The practical response for defenders: assume commodity attacks now arrive in greater volume and greater variety, tighten detection for AI-generated malware families that mutate faster than signature databases update, and watch for agentic tooling (Claude Code, open-source equivalents) executing post-breach commands. If your EDR flags a coding agent spinning up inside a production network you did not authorize, that is no longer a hypothetical.

The AI industry has spent two years debating whether models could meaningfully assist attackers. TeamT5’s data suggests the debate is over. They already do, at scale, and the doubling has happened.

Sources

• Bloomberg
• Implicator.ai