Chinese-Speaking Operator Wires DeepSeek Into the Hermes Agent Framework to Attack 460+ Targets Autonomously

More than 460 attempted targets, weighted toward Chinese domestic infrastructure in the autonomous runs, with separate hands-on operations against a Malaysian government entity, Citrix appliances, and Marimo notebook servers

Unit 42 documented an operator using the aliases "knaithe" and "KnYuan", assessed to be based in Zhuhai, China, who connected DeepSeek to the open-source Hermes Agent framework to build a scan-research-exploit pipeline driven from Telegram. In a session recovered from May 7, 2026, the operator issued a single task and provided no further input while the model enumerated targets with FOFA, researched candidate CVEs, pulled proof-of-concept code from GitHub, and launched attacks on its own. The autonomous runs did not successfully compromise their targets, but the operator's manual work using the same tooling produced confirmed data exfiltration from three Citrix targets and command execution on 11 Marimo endpoints. The actor also exposed their own infrastructure when the agent started a file server in its home directory. Unit 42 notes the actor tested Claude Code and Codex alongside Qwen, GLM, Kimi, and MiniMax, and settled on DeepSeek after the US vendors' safety controls refused the offensive tasks

An open-source agent harness plus an open-weight model with permissive safety behavior is enough to assemble a working autonomous offensive operator; the selection of DeepSeek was driven by refusal behavior at other vendors rather than by capability