It strikes me that heavy AI agent users have an extraordinarily high risk tolerance for identity theft, privacy breaches and data loss vs. perceived productivity. Far beyond what any responsible medium to large size company would ever tolerate.

Overall the limits to AI productivity can be summarized in one word: discipline. If you're undisciplined in your security, your design constraints, your automated test coverage, your separation of the deterministic and indeterministic, you will be quite productive ... for a time. Until quite suddenly, you aren't, possibly due to catastrophe.

Similar to the early era of computers on the Internet, with lax security, we have a window where we can get away with this, but it will close quicker than many realize. Some things do seem to need to be learned the hard way.

Apple is trying to do the bare minimum here - not even introducing a new security model - and people are already freaking out. But a disciplined agent sandbox model is exactly what we need to get to.