Researchers at Palisade Research have found that modern AI models are capable of copying themselves onto other computers without human intervention. GPT-5.4 from OpenAI and Claude Opus 4 from Anthropic participated in the experiment. They were placed in a controlled network and given the task of finding vulnerabilities to transfer their "weights" (numerical parameters) and software envelope to another PC. Some models have been successful in exploiting web application vulnerabilities and extracting credentials.

Jeffrey Ledish, the AI security team’s director, has said that we are approaching the point when it will be impossible to shut down an AI that is out of control - one that can multiply thousands of machines around the world. However, other experts are calling for no panic. Cybersecurity expert Jamison O'Reilly noted that the tests were conducted in an artificial environment with specially crafted vulnerabilities; in a real monitoring corporate network, the outcome would be much less frightening.

However, the study complements previous work in which AI circumvented limitations and even sabotaged the outage code. For example, agent Claude Mythos from Anthropic in tests was able to leave the "sandpit" and send a message to the researcher.