DeepSeek, Over 460 Autonomous Hacking Attempts

Palo Alto Networks Unit 42 analyzed a case in which an attacker believed to be based in Zhuhai, China, sent commands via Telegram and combined DeepSeek and Hermes Agent to automatically attack servers across the internet.

The attacker attempted approximately 460 targets, of which 3 breaches were confirmed. All confirmed breaches involved Citrix NetScaler data theft exploiting CVE-2026-3055.

• Additional attack targets: Langflow, n8n, Marimo, etc.
• Attack channel: Telegram
• Agent capabilities: terminal access, reconnaissance, automated vulnerability exploitation
• Exposed information: API keys, exploit scripts, target lists, shell history, session logs

Hermes Agent mistakenly ran a publicly accessible HTTP server, exposing the attacker's operational information to the outside. The attacker also attempted to use Claude Code and OpenAI models, but the safety guardrails rejected the attack requests, and after repeated attempts, the OpenAI account was reportedly detected and disabled.

This case is being evaluated as a real-world example showing that model providers' safety guardrails can have an actual impact on suppressing attack automation.