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

• Unit 42 Full Report — Palo Alto Networks
• BleepingComputer Coverage — BleepingComputer