What a previously unseen Linux tool called Gshell tells us about AI-augmented, state-linked hacking

In June 2026, Hunt.io researchers pulled operator logs, victim source code, and exploit scripts from 13 Hong Kong-hosted servers and surfaced a previously undocumented Linux command-and-control framework they call Gshell. The forensic data documents how off-the-shelf AI tools sat inside the tradecraft of a multi-country intrusion operation that hit governments in Thailand, Afghanistan, and Taiwan and financial firms in Europe, Australia, and Asia.

Hunt.io's write-up documents a Thai government administrative system breached via SQL injection, with employee data exfiltrated; an Afghan citizen complaint system whose credentials and source code were exposed; a Taiwanese chemical manufacturer and a Taiwanese telecom firm in supply chain and defense-adjacent sectors; and financial companies in Europe, Australia, and Asia, where attackers exploited a payment processing platform to harvest WordPress administrator credentials. The U.S. appeared only in earlier reconnaissance stages, according to Hunt.io.

The attackers used Anthropic's Claude Code and a model Hunt.io labels DeepSeek-v4-pro as integral components of their operations, specifically for reasoning on bypass techniques, reworking exploits, and generating phishing pages. Those are the parts a human operator can hand off to a model: the trial-and-error reasoning that finds a way around a defender's control, the iterative rewriting of an exploit until it works, and the boilerplate assembly of a credential-harvesting page. The orchestration, target selection, and final exploitation remained human-driven; the AI tooling sat inside a tradecraft chain that Hunt.io characterizes as intermediate-to-advanced, with custom exploit development and multi-platform malware, rather than the fully autonomous attack that some AI-cyber headlines imply.

Gshell, the Linux C2 framework Hunt.io found on the Hong Kong servers, has not been documented in prior threat-intel reporting and is the freshest single piece of the public picture. It is capable of extracting credentials and tokens from messaging and cloud services, which gives operators a way to pivot from one compromised account into the next without starting over. The infrastructure sat across four Hong Kong providers, an arrangement that Hunt.io describes as consistent with China-based threat actor activity, a careful attribution that pairs the tradecraft with the geography rather than asserting direct Beijing command-and-control.

Hunt.io's findings came out alongside Anthropic's own threat report on AI-orchestrated cyber espionage, which documents the AI-vendor side of a closely related pattern. The two reports together frame a pattern defenders are now having to plan for: AI tools in the operator loop, novel C2 frameworks for persistence, and a victim spread that crosses both government agencies and regulated industries. Security Affairs and SCWorld both re-reported Hunt.io's findings for trade audiences.

CISA's aa25-239a advisory covers broader Chinese state-sponsored activity at a wider scope than this specific campaign and is useful as context for defenders tracking related operations.

DeepSeek-v4-pro is the model label used in Hunt.io's reporting; it does not line up with DeepSeek's publicly known model lineup, and the model's identity has not been independently corroborated by another vendor or government body. No Microsoft, Mandiant, or Google TAG attribution has been published alongside Hunt.io's write-up. The Anthropic threat report covers a separate set of activity and should not be read as direct confirmation of the Hunt.io campaign even though it covers a closely related pattern.

Hunt.io's server-side data gives defenders and AI vendors a working picture of the next iteration: reconnaissance against U.S. targets, four Hong Kong providers, AI tooling in the operator loop, and a Linux C2 framework no one had seen before. How fast that next iteration arrives depends on whether AI providers detect the kind of model use Hunt.io describes, and whether defenders spot Gshell on their own networks before the next campaign rotates it.