Hackers use AI agent tools such as ARTEX in attacks
One hacker group went from a first account to deep system access in four hours.
Hackers use open-source AI agent tools in attacks. The threat team of Google and its Mandiant unit sees this since August 2026. The tools are ARTEX, Hexstrike, PentAGI and CAIRN, and personnel who test networks made them. Google sees that the agents try many attack paths, correct errors, and do goals in seconds or minutes.
How it works
ARTEX divides an attack into tasks for six AI helpers. It uses models to change a goal into commands. Google saw it show user accounts, get more rights, find cloud storage keys, and write webshells (small tools for remote control). Hackers gave it logins of other users, and it did the next steps.
The numbers
The group that Google names UNC7228 hit firms in the US and Europe. Wiz, a cloud firm, found it. Its access time was about four hours, and one time 15 minutes. Google gave no count of all attacks.
What is not known
The maker of ARTEX removed the code in October 2026 because of abuse. Copies continue to move, with ScopeWeaver and a commercial version, ARTEX PRO. Google sees that the methods are the same as before: logins of other users and systems with no patches. Only the speed and scale are new.
Insight: The code is free to copy but model access is not
The maker removed ARTEX, but copies continue. Zscaler reports that one operator used it with Claude Code, Codex and Grok. Session records of Claude Code helped CrowdStrike connect the attacks on banks in South Korea. The code is easy to copy, but model accounts are not. The primary position to monitor this reads as the model provider. Look for more requests for their logs.
Sources
Posted
