The dangerous myth behind AI agent hacks

These are not simply cyber security flaws but problems with the way that models are trained