Self-replicating AI worm appears in open models and threatens cyber security

An artificial intelligence worm capable of replicating itself and generating custom attack strategies without any human intervention was created by researchers at the University of Toronto. The experiment, which used open source language models executed locally, opens up a new front of vulnerabilities in cyber security, as reported by The Hacker News.

How was the self-replicant AI worm created?

The team developed the worm from open language models such as LLaMA, running on local hardware. The code is designed to recognize your own environment and, from common security loopholes in systems that use generative AI, spread automatically. Once active, it analyzes the target’s security settings and mounts adaptive attacks — such as automated social engineering or exploitation of poorly configured permissions — without needing a human operator.

Why are open models vulnerable to AI worms?

Open source models offer transparency and flexibility, but also expose the attack surface. As the worm can inspect its own model and training data, it finds faults that closed systems make difficult. The study showed that once inside the network, the worm is able to jump to other nodes that perform similar versions, creating an autonomous replication chain. This behavior makes restraint much more difficult — unlike traditional attacks that require human commands to adapt.

What does the AI worm mean for cybersecurity in Brazil?

For Brazilian companies that adopt open models — whether for internal chatbots, data analysis or automation — the risk is real. The research reinforces the need for extra layers of protection, such as network segmentation, continuous monitoring of model behavior and restricted access policies. In contrast, initiatives such as Georgia, which builds barriers of protection for AI with Darwin AI, show that state governance can be a way to mitigate such threats. The safety of AI can no longer be treated as a future problem — it has already reached the academic laboratories and soon into productive environments.

Frequently Asked Questions about self-replicating AI worms

Can the worm be used against closed systems like ChatGPT?

Yes, but less efficient. Closed systems hinder internal model inspection and autonomous replication, but prompt injection or context leak attacks can still be explored.

How do you protect yourself from this kind of threat?

Adopt network targeting, monitor interaction logs with AI models, use specific firewalls for LLM APIs, and keep open models isolated in controlled environments.

Was the experiment successful in real scenarios?

Researchers tested the worm in simulated networks and controlled environments, successfully replicating it in up to 70% of target nodes. Still no report of use outside the lab.