Researchers demonstrate how AI may be vulnerable to viruses

Security researchers at Cornell University in the US have demonstrated the effectiveness – and therefore the potential danger – of a virus capable of infiltrating tools powered by artificial intelligence, stealing valuable information and even sending malicious emails.

The researchers focused on the use of a computer worm – a self-replicating type of malware – targeting applications powered by generative artificial intelligence. In the case of generative AI-powered email assistants, for example, the worm was able to exfiltrate personal data and spam numerous users.

The scientists were thus able to demonstrate that a dedicated worm can spread rapidly and automatically between different generative AI agents. Once infiltrated, the worm uses adversarial self-replicating prompts designed to produce more harmful prompts and capture data that is supposed to be private.

Unfortunately, this experiment foreshadows a new type of cyberattack. Based on their research, the scientists predict the arrival of this type of worm within the next two to three years at most. And the growing number of applications available based on generative AI only increases the risk of this type of threat spreading.

However, a classic data security approach should limit the risks, according to the researchers’ initial observations. The idea is that humans should always be involved in certain authorisations, so that no data is able to leak out automatically.

Symbolically, the researchers have named their virus Morris II in a nod to the famous Morris worm, a pioneer of experimental viruses on the Internet, launched in 1988. They have shared the findings of their research online to raise awareness of this new type of attack. – AFP Relaxnews

Tagged