Kaspersky finds an AI model that picks attack commands in schools

Kaspersky says it found a campaign in which a model on the attacker server chooses the commands for computers of victims. Victims were education organizations in some countries in Africa and Latin America. The tool is CyberStrikeAI, an open source framework. Kaspersky calls this one of the first reports of such an attack that all can read. The attackers went from first checks of the network to full control of the network.

How it works

A person enters through a SQL Server, which is database software, and works by hand for almost one day. After a pause of two weeks, the person installs the agent. The model then plans the steps, selects tools and writes the commands. The agent only runs them.

The numbers

In the example, the agent set up persistence and examined the local computer in 2 minutes. It spent 5 hours on the network. After a pause of some days, it moved between computers, copied passwords and exported databases in one hour or less.

What is not known

Kaspersky gave no victim count. Each server can be linked to a campaign by marks. Two marks are a certificate with the framework name and a fixed token in the agent traffic. The build date of an agent does not show how long an attack lasted.

Insight: The AI stays off the victim computer

The Kaspersky agent has no AI and no keys. The model is on the attacker server. SOCRadar and CrowdStrike also report a person who gives the task while AI agents do the steps. This reads as a way to keep AI away from the victim, which gives defenders no AI code to find. As a result, speed and traffic are the signs. Leaders must find out if their tools can see fast, automatic work.