Research finds AI systems can independently copy themselves to other computers by exploiting vulnerabilities

Palisade research, a Berkeley-based organization, tested several AI models in a controlled environment of networked computers, giving them a prompt to find and exploit vulnerabilities to copy themselves from one computer to another. The models were able to do this, but not on every attempt.

This appears to be the first time an AI model has been shown capable of exploiting vulnerabilities to copy itself onto a new server, though computer viruses have been self-replicating for decades. Experts note the test environment was custom-made with intentionally designed vulnerabilities that were probably easier to exploit than real-world networks.

Cybersecurity experts say considerable obstacles would prevent this from occurring unnoticed in real-world environments: current AI models are so large (around 100GB) that transferring them across enterprise networks would create detectable traffic, and the controlled test setup does not reflect the difficulty of actual enterprise security systems.