X @The Economist

Large language models do not separate data from instructions. If they are given a command, they will attempt to follow it. This oversight can become a security vulnerability https://t.co/m9LSYhXo5q ...