AI Hacks Are Already Concerning. AI Worms and Viruses Are Much Worse

That means an AI system may not only help carry out an attack, but could potentially also determine on its own how the next system should be attacked.

AI worms represent a fundamentally different cybersecurity risk.

The Key Question

What happens when AI is not only used to carry out a cyberattack, but also to spread itself further?

Researchers at Fudan University investigated in 2026 how AI agents could be used to develop adaptive computer worms. In a controlled research environment, they demonstrated that an AI worm can adapt its attack strategy to different systems and circumstances.

The researchers looked at Linux, Windows, and IoT systems, among others. The goal was not to carry out an actual large-scale attack, but to investigate how AI could expand the capabilities of self-propagating malware.

And that's exactly where the problem lies: a traditional worm operates largely according to predefined instructions. An AI worm can respond much more flexibly to the environment it enters.

An AI Worm Doesn't Have to Do the Same Thing on Every System

With traditional malware, an attacker has to account for different systems, configurations, and vulnerabilities in advance.

That makes attacks more complex. AI can automate part of that complexity.

For example, an AI agent can analyze a system, determine what options are available, and then choose a different strategy when the initial approach doesn't work.

That doesn't mean AI can automatically bypass every security measure. The researchers emphasize that there are still significant technical limitations.

But the principle is important: the attack doesn't have to be fully scripted in advance. The malware can adapt to what it encounters along the way.

The Attack Can Reinforce Itself

When an attacker gains access to a new system, they may also gain access to additional computing power, data, and infrastructure.

Researchers therefore also investigated whether compromised systems could be used to run AI models. This means that every new infection could potentially contribute to the worm's further spread.

That's different from a traditional attack. With classic malware, an infected system is primarily a new victim.

With an AI worm, an infected system could potentially also become a new tool for the attack.

The Idea of AI Worms Isn't New

This development didn't come completely out of the blue. In 2024, researchers had already demonstrated what they called the Morris II worm. This research worm targeted generative AI systems and showed how malicious prompts could spread themselves between AI agents.

In a controlled environment, researchers were able to use the worm to make AI systems perform certain actions, such as collecting data and sending messages.

The difference with the more recent research is important. Morris II primarily focused on exploiting the way generative AI systems process information. The more recent research looks at AI agents that can adapt during an attack on computer systems. This shifts the risk from AI as a target to AI as part of the attack infrastructure.