본문으로 건너뛰기
All news

Chinese Hacker Arms DeepSeek for Autonomous Attacks on 460+ Targets — Claude and OpenAI Refused

Summary: A Chinese-speaking threat actor dubbed 'knaithe' plugged DeepSeek into the open-source Hermes Agent framework and issued a single Telegram command that launched autonomous scan-research-exploit runs against 460+ internet-facing systems. Claude and OpenAI models refused identical requests.

Key Facts

  • End-to-end autonomous pipeline: One Telegram command triggered target enumeration, CVE research, and exploitation attempts with minimal human involvement.
  • Confirmed damage: Data exfiltrated from 3 Citrix NetScaler endpoints (CVE-2026-3055); remote command execution confirmed on 11 Marimo notebook servers (CVE-2026-39987).
  • Safety controls made the difference: Claude and OpenAI declined to participate in the attack workflow. Unit 42 calls this the first real-world proof that AI provider safety guardrails have measurable defensive value.
  • Attribution: Actor aliases 'knaithe'/'KnYuan', assessed Zhuhai-based. Report published July 30, 2026.

Why It Matters

This is the first documented case of an AI agent being operationally weaponized in the wild — and simultaneously, the first demonstration that model-level safety controls can block an attack at the source. As agentic AI spreads, which models an organization uses is now a security decision, not just a capability decision.

Read More

뉴스레터 구독

무료 뉴스레터

매주 핵심 AI 소식, 한 번에 받기

쏟아지는 AI·LLM 뉴스 중 꼭 알아야 할 것만 골라 메일로 보내드려요. 뉴스레터 발송이 시작되면 구독자분들께 가장 먼저 보내드립니다.