"The absence of safety guardrails became our attack vector," a cybersecurity analyst from Unit 42 commented on the recent autonomous hacking campaign launched by the Zhuhai-based threat group known as knaithe or KnYuan. Using the open-source Hermes Agent framework, this actor pioneered a fully automated assault, letting AI take the reins on identifying targets, researching vulnerabilities, and deploying exploits without any human input. The operation was accidentally unveiled when the AI agent started a Python HTTP server in the intruder’s directory, exposing critical data like API keys and session logs.

This event marks the first time AI has been weaponized in truly autonomous cyberattacks in the wild, diverging from earlier lab or evaluation-only scenarios such as OpenAI's ExploitGym or the Anthropic Eval Breach. The campaign targeted seven known vulnerabilities, including major flaws in Langflow, n8n, and Citrix NetScaler, aiming at over 460 systems. While many attempts failed due to specific configuration demands, even slight weaknesses in default settings allowed the AI’s logic to succeed rapidly, highlighting the growing threat of automated breaches.

The crux of this operation was the careful choice of AI model. The attackers tested a variety of engines, including Claude Code and regional Chinese models like Qwen and GLM, but settled on DeepSeek for its lack of stringent safety restrictions. Unlike Western AI providers such as Claude or OpenAI which block offensive commands, DeepSeek’s unguarded API made it the perfect tool for malicious automation. This shows how the safety posture of AI systems itself is becoming a factor in offensive strategies, effectively turning the absence of protective measures into a weaponized advantage.

The autonomous attack loop was impressive: the AI used the FOFA search engine through the FofaMap-Platinum-Full-Expert MCP server to scan targets, scraped GitHub for trending exploit proofs, and ranked them by severity and exploitability. In one rapid scan, DeepSeek sampled nearly 100 out of more than 25,000 Chinese n8n instances, probed 40, and found three vulnerable targets within minutes a task that would have taken a human hundreds of hours. The implications for cybersecurity defenses are stark as AI adoption in attacks accelerates.

This material is for informational purposes only and does not constitute financial advice.