But I wouldn’t run my production apps—that actually make money or could cause harm if they break—on unreviewed AI code.

Here’s the thing though… if you design your system and security boundaries well (which isn’t always easy), you can run non- or poorly reviewed code relatively safely. It’s standard security practice to design for minimal privilege.

What you don’t ever want to do is give AI agents - nor inexperienced engineers - admin permissions and let them run wild. Both AI and human will do something stupid at some point, leaving you at risk.

For serious companies, the risks are more likely to be on the developer’s own machine, when they give it admin credentials and unsupervised access. For example, when “Fix this bug” leads to it testing in production, because it has both prod credentials and configuration at its disposal.

Lesson: Give agents and humans safe playgrounds that limit the blast radius, both during development and in production.