A Chinese threat actor has reportedly leveraged large language models (LLMs), including DeepSeek AI's Hermes Agent, to automate and accelerate vulnerability exploitation campaigns against internet-exposed infrastructure in Asia. The actor, known by the aliases 'knaithe' and 'KnYuan' and believed to be based in Zhuhai, China, is described as an opportunistic exploit operator and binary security researcher.
According to a report published on July 30 by Unit 42, Palo Alto Networks' research team, the actor utilized Hermes Agent, an open-source agentic AI framework connected to a DeepSeek AI model, to orchestrate attacks via Telegram. This AI-augmented approach significantly increased the speed and scale of their offensive operations.
The campaign involved a combination of autonomous AI-driven enumeration and automated exploitation, alongside manual exploitation of several vulnerabilities. When initial automated exploitation attempts failed due to restrictive target configurations, the Hermes Agent autonomously searched for critical-severity Common Vulnerabilities and Exposures (CVEs). It initially surveyed ten product families, scanning GitHub for trending proofs of concept (PoCs) and prioritizing vulnerabilities based on attack surface.
This research led the agent to pivot to seven specific vulnerabilities. Autonomous exploitation was attempted for CVE-2026-33017 (Langflow, CVSS 9.8) and two n8n Workflow Automation vulnerabilities, CVE-2026-21858 (CVSS 10.0) and CVE-2025-68613 (CVSS 9.9), though these attempts failed due to disabled auto-login or authentication requirements.
Manual active exploitation was observed for CVE-2026-3055 (Citrix NetScaler ADC & Gateway, CVSS 9.8), resulting in data exfiltration. Reverse shell attempts were made against CVE-2026-34486 (Apache Tomcat, CVSS 7.5) and CVE-2026-33824 (Windows IKE Extensions, CVSS 9.8). Command execution was confirmed through manual active exploitation of CVE-2026-39987 (Marimo Notebook, CVSS 9.8). Additionally, a non-functional research PoC for CVE-2026-0300 (PAN-OS User-ID Authentication Portal, CVSS 9.8) was cloned but not executed. The autonomous attack flow was observed in a Hermes Agent session on May 5, 2026.
In parallel with DeepSeek, the actor configured and tested multiple other LLMs, including Chinese models like Qwen, GLM, Kimi, and MiniMax. Limited use and testing of Western AI tools, such as Claude Code for connectivity and proxy validation, and OpenAI’s Codex for exploit development, were also noted. This suggests an evaluation phase to identify preferred AI tools.
While the observed campaign had limited impact and did not achieve full compromise of all intended targets, it demonstrated a functional, end-to-end autonomous offensive capability. Targets spanned three countries, including China and Malaysia, and multiple sectors.
Analysts emphasize that the significance lies in the trajectory of these operations rather than the immediate outcome. The actor is actively refining tool configurations, developing custom skills, establishing proxy infrastructure, and executing autonomous attack cycles. This indicates a decreasing technical barrier to AI-augmented offensive operations.
