AI agents are already carrying out cyber attacks with almost no human intervention, writing programs on their own, looking for penetration routes and correcting their own mistakes. Wiz researchers analyzed three real-life hacks and found telltale signs of machine work that help distinguish them from the actions of ordinary hackers.

In the first case, the attackers penetrated the Internet-accessible server of Flowise, a platform for creating AI applications. By bypassing the access check, the attackers were able to run system commands. The hacked server was then turned into a platform for attacks on dozens of other Flowise instances.

Within 24 hours, the attackers created 99 different scripts. New programs appeared every one or two minutes, checked the availability of targets and selected further actions. The code retained detailed plans in natural language, and one program directly explained why the previous version was not validating API keys correctly and how to fix the error.

Wiz observed the second attack after a corporate gateway in Java was hacked. In four and a half hours, the attackers completed 602 unique steps, testing 12 attack vectors. Among the targets were Kubernetes, RabbitMQ, ArgoCD and internal services. Failed requests did not stop the process: the tools changed tactics and tried other routes.

In the third case, the entry point was the GraphQL service of an AI agent with an MCP interface. The attackers succeeded in executing commands, stole cloud credentials and Kubernetes tokens, gained a foothold in the system, and spread the attack to multiple containers. In 21 hours, more than 150 scripts appeared, and the found files contained 803 lines of explanatory comments.

Particularly revealing was the recording in which the performer compared his own results with the work of “another agent.” The comments discussed unsuccessful attempts to access the database and the reasons for the differences between launches. Such a trace does not prove the complete autonomy of the operation, but directly indicates the participation of the agent system in technical actions.

Wiz has identified seven signs of such attacks. These include continuous work at a variable pace, detailed explanations in scripts, self-correction of errors, trying out many hacking methods and a large number of temporary files. Such actions complement the already known scenarios of agent attacks using trusted tools.

The review also cites cases with ARTEX and Hermes instruments. Specifically, Storm-0940 used ARTEX to accelerate attacks against ConnectWise Cloud Backup clients. The Wiz study emphasizes that the level of autonomy varies across episodes, and some of the findings are based on indirect evidence.

Speed does not always mean stealth. Agents go through dozens of options, create redundant files, and leave detailed activity logs. Defenders must connect disparate events across cloud systems and block suspicious activity faster. At the same time, automation of attacks already allows attackers to simultaneously check many more targets.