IC Technology & Cyber — 2026-07-18
Hunt.io Discovers Chinese Cyber Espionage Campaign Using Claude Code and DeepSeek AI to Automate Attacks on Government Systems Across Four Countries
BLUFConfirmed operational use of dual-model AI automation in Chinese intrusion sets marks the shift from experimental tooling to standardized tradecraft, outpacing current vendor abuse-detection architectures.
Hunt.io researchers reported that a pivot from known TencShell command-and-control infrastructure led them in June 2026 to an open directory on a Hong Kong-based server containing 2,431 files and 80 subdirectories, including victim source code, exploit scripts, and operator logs written in Simplified Chinese 12. Hunt.io assessed that Claude Code handled execution, including bash commands and phishing-page construction, while DeepSeek-v4-pro drove attack reasoning and exploit revisions, with operator sessions dated June 8 through 12, 2026 12. The researchers documented active exploitation of government systems in Thailand, where SQL injection gave attackers admin access and exposed employee national ID data, and in Afghanistan, where a citizen complaint portal built on Laravel was breached for source code and credentials 12. In Taiwan, Hunt.io identified reconnaissance against eight organizations with two, a chemical manufacturer and a telecom and edge-device maker, successfully exploited, while a parallel campaign extracted WordPress administrator credentials from a payment processing platform serving Europe, Australia, and Asia 1. A fourth country, the United States, saw reconnaissance-only activity with cloned phishing pages built for the D.C. Council and Delaware County government sites, part of broader scanning that touched more than 5,890 government hosts across 10 countries. Hunt.io traced the infrastructure to 13 Hong Kong servers across four hosting providers, identified an undocumented C2 framework it named Gshell, and notified affected organizations and national CERTs on July 6 ahead of publication 123.
Analysis
Hunt.io's find, resting on its own single-source research republished without independent corroboration, confirms agentic AI has moved from proof-of-concept to sustained use in Chinese state-linked intrusions, with Claude Code handling execution and DeepSeek-v4-pro driving exploit iteration across four countries' government and financial networks. Pairing US-origin agentic execution with a domestic Chinese reasoning model complicates attribution and undercuts vendor detection built around single-model authorship. Confirmed compromises in Thailand and Afghanistan, plus staged financial-sector credential theft, suggest operators retain live access despite the July 6 CERT notifications, and the recurrence after Anthropic's November 2025 disclosure marks this as repeatable tradecraft rather than an isolated experiment. The underlying vectors, SQL injection, hardcoded keys, Laravel flaws, remain conventional, so the AI layer may be accelerating existing tradecraft rather than expanding offensive capability.
3 sources
- Suspected Chinese Operators Use Claude Code and DeepSeek to Target Government and Financial Systems Across Four Countries - Hunt.io
- Claude Code and DeepSeek Powered Chinese Cyber Espionage Campaign - Security Affairs
- Chinese actors leverage AI tools for automated cyberattacks on government and financial systems - SC Media