Chinese Hacker Uses DeepSeek via Telegram to Drive Autonomous Attacks

Palo Alto Networks’ Unit 42 has uncovered a Chinese-speaking threat actor who leveraged the DeepSeek AI model through the open-source Hermes Agent framework to conduct autonomous cyberattacks. The operator, tracked under aliases knaithe and KnYuan, initiated the operation with a single Telegram instruction, after which the agent independently identified internet-facing systems, selected public exploits, and launched attacks against more than 460 targets.

Unit 42 detailed seven exploit tracks spanning eight CVE identifiers, including attempts against Langflow, n8n, NetScaler, and Marimo. The DeepSeek-led attacks on Langflow and n8n failed due to configuration mismatches, but manual operations reportedly led to data exfiltration from three organizations via CVE-2026-3055 (NetScaler memory-overread) and command execution on 11 Marimo instances via CVE-2026-39987. However, the report later states only three targets were successfully exploited, a discrepancy that The Hacker News has asked Palo Alto Networks to clarify.

The agent’s behavior included checking versions, downloading exploits, abandoning unproductive paths, and selecting vulnerabilities based on severity, deployment scale, and exploitability. The operation was exposed when Hermes Agent inadvertently started an HTTP server, revealing model configurations, API keys, exploit scripts, target lists, and session logs. DeepSeek was the primary reasoning model, with limited use of Claude Code and Qwen Code, and possible but unverified use of Codex.

Organizations are advised to patch exposed Langflow, n8n, and Marimo systems, as well as customer-managed NetScaler ADC/Gateway appliances configured as SAML identity providers. Removing unnecessary public access to workflow and notebook interfaces is also recommended. The operator is assessed to be based in Zhuhai, China, though this is not independently verified.