Chinese Hackers Automate Cyberattacks Using DeepSeek AI
DeepSeek AI autonomously manages target scanning and attack initiation processes.
1Discovery of Autonomous Attack Chain
Palo Alto's Unit 42 team uncovered Chinese hackers using DeepSeek AI integrated with Hermes Agent framework to launch fully autonomous attacks. The system independently executed target scanning, vulnerability searching, and attack initiation processes. Researchers obtained this evidence after hackers accidentally exposed operational data by misconfiguring their server directory.
2Multiple AI Experimentation
Attackers tested Chinese-origin language models including Qwen, GLM, Kimi, and MiniMax alongside DeepSeek. Western tools Claude Code and Codex were used minimally. Notably, Western tools were routed through proxies while Chinese models connected directly to native APIs, clearly showing which tools the attackers trusted for actual operations.
3Langflow Vulnerability Attempt
DeepSeek identified CVE-2026-33017 in Langflow and scanned 84 live instances. All targets required either public flow IDs or inaccessible login settings. The AI analyzed the small target base and decided to search for larger targets, leading it to n8n workflow automation tool with over 647,000 global instances.
4Autonomous n8n Exploration
DeepSeek detected over 25,000 China-based n8n instances. It chained two vulnerabilities to identify exploitable version ranges. The system sampled 100 IP addresses, actively probed 40, and found 3 vulnerable systems - compressing what would normally take human operators hours into minutes according to Unit 42.
5Successful Manual Attacks
While the autonomous system failed, the same actor successfully exfiltrated data from 3 Citrix NetScaler targets using traditional methods. They achieved command execution on 11 Marimo notebook endpoints. In multi-day attacks against a Malaysian government body, they progressively increased anonymization techniques.
6Red Team Customizations
Hackers customized Hermes Agent with three specialized red team skills: FOFA cyberspace search for asset enumeration, LLM jailbreaking capabilities, and framework-bundled WebSocket exploitation tools. These customizations enabled the AI to conduct more sophisticated autonomous operations while evading detection.
7Industry Impact
This incident demonstrates AI's rapidly evolving role in cybersecurity. Unit 42's report suggests autonomous systems can bypass traditional defenses. The preference for Chinese-origin AI tools in cyber operations creates new national security concerns. Industry experts predict a 300% increase in AI-powered attacks within 5 years.
8Defense Strategies
The report warns traditional security measures may be inadequate against autonomous attacks, urging development of AI-powered threat hunting systems. It highlights urgent needs in API security and anomaly detection. Unit 42 recommends companies test defenses using AI-powered attack simulations to prepare for this new threat landscape.