Kaspersky has described one of the first publicly documented cases of a large language model sitting inside the core of a real attack, after researchers found an AI-directed hacking tool driving intrusions into school networks in Africa and Latin America.

The tool, CyberStrikeAI, is an open-source penetration-testing framework written in Go. Its language-model orchestrator runs on the attackers' command-and-control server, where it plans actions, picks tools and writes the commands, while a small implant on each infected machine simply carries them out. The implant itself contains no AI, no service keys and makes no decisions of its own, according to Kaspersky's Securelist report, published on 2 November 2026.

The confirmed victims were education-sector organisations in several countries across Africa and Latin America. A human operator first broke in through a compromised Microsoft SQL Server, spent about a day gathering information and exploring the network, then went quiet for around two weeks before deploying the implant.

From there the AI-driven phase moved fast: the attackers reached full control of the victims' Windows domains, the central systems that manage users and computers, within three to four days. In the final sequence, lateral movement, mass credential theft and database export took under an hour, Kaspersky said. One implant was disguised as “svchosts.exe”, mimicking the legitimate Windows process, and persistence was set through a scheduled task named “SystemUpdate”.

Kaspersky did not name a threat group or say how many organisations were hit, and it did not quantify any stolen data. Chinese-language comments in the scripts, use of a Chinese DNS-logging service and server restarts during UTC+8 business hours point to a Chinese-speaking operator, the researchers said, an assessment rather than a confirmed attribution. The researchers added that the compromised hosts still showed familiar attack behaviour that existing detection tools can catch.