IncidentAI-enabled attacks
An open-source agent and DeepSeek run hundreds of attacks for one operator
HighPalo Alto Networks Unit 42
Unit 42 documented a single operator who used the open-source Hermes Agent framework with DeepSeek as its reasoning model to find internet-facing systems, pick public exploits and attack them with little further input. It went after more than 460 targets; Unit 42 confirmed data theft from three organizations and command execution on 11 servers.
Replay the attack step by step in the Learning Center
How it happened
- After one instruction over Telegram, the agent enumerated exposed services such as Langflow, n8n, Citrix NetScaler and Marimo notebooks.
- It chose matching public exploits and ran them, retrying and adapting on its own.
- The operator also tried other models and coding agents, including Claude Code, Codex and several Chinese LLMs.
- The agent itself exposed the operation: it started a web server in its home directory that published logs and API keys.
Impact
Data was exfiltrated from three Citrix NetScaler targets and commands ran on 11 Marimo instances; most attempts failed on patched or well-configured targets.
The fix
Unit 42 published the targeted CVEs and indicators; the flaws exploited had fixes available.
Lessons
- One person with an agent now scans and exploits like a team. Patch exposed AI and edge software on days, not months.
- Hardened configuration defeated most of the automated attempts.
- Keep AI platforms such as Langflow and n8n off the internet.