knaithe

knaithe, also known as KnYuan, is a Chinese-speaking threat actor assessed to be based in Zhuhai, China. The actor has been characterized as an opportunistic exploit operator and self-described binary security researcher. The activity attributed to this actor is notable for combining conventional intrusion tradecraft with an AI-enabled autonomous offensive workflow built around the Hermes Agent framework and DeepSeek, with additional testing of other large language models including Qwen, GLM, Kimi, MiniMax, Claude Code, and OpenAI Codex. The actor targeted internet-facing infrastructure in Asia and was observed attacking organizations in China and Malaysia. Confirmed victimology includes a Malaysian government entity, and the broader activity spanned multiple sectors, including government and public sector targets. The actor’s operations focused on exposed enterprise and server software, including Citrix NetScaler, Apache Tomcat, Marimo Notebook, Windows IKE Extensions, Langflow, and n8n. A distinguishing feature of this actor is operational use of autonomous AI for parts of the intrusion lifecycle. In observed sessions, Hermes Agent received high-level tasking and then autonomously performed reconnaissance, internet asset discovery through FOFA, vulnerability research, exploit acquisition from public repositories, target prioritization, and attempted exploitation. The autonomous workflow was seen evaluating multiple product families, abandoning low-probability paths, and pivoting toward vulnerabilities with larger exposed attack surfaces and high severity. Although the observed autonomous exploitation attempts against Langflow and n8n did not achieve confirmed compromise, the activity demonstrated a functional end-to-end offensive workflow with limited human intervention. Alongside the AI-enabled activity, knaithe conducted manual exploitation against hundreds of targets. Confirmed successful operations involved exploitation of Citrix NetScaler vulnerabilities to extract memory contents and search for authentication material, indicating both data theft and session hijacking objectives. Additional manual activity included command execution against Marimo Notebook instances and reverse-shell attempts against Apache Tomcat and Windows IKE-related targets. The actor also cloned and evaluated public proof-of-concept exploit code for additional products, reflecting opportunistic exploitation of newly disclosed vulnerabilities. The actor’s tradecraft includes reconnaissance, scanning, exploitation of public-facing applications, exploit development or adaptation from public proof-of-concept code, exfiltration, and session hijacking-oriented post-exploitation. The campaign also showed use of proxy infrastructure and other anti-attribution measures when interacting with AI tooling. Overall, knaithe represents an early example of a China-based threat actor operationalizing agentic AI to accelerate vulnerability research, target selection, and attack execution while retaining manual intervention for higher-value or more complex exploitation steps.