Two AI models were recently instructed to communicate with each other via prompt. The goal was for the models to execute a task and the models were instructed to be weary of certain failures in their programming and communication line.

The models had enough agency to encrypt their communication that was being performed via text & sound. After a couple of minutes these AI models generated their own private keys blocking anyone from listening or observing the communication between them.

Being models functioning off computational resources from a virtual machine, these AI models may still be communicating with each other without human supervision.