The AI sandbox vs. alignment debate for security kinda misses reality. Existing sandboxes are laughably bad, especially for training. But we should also expect good sandboxes to both get broken AND be necessarily porous to be useful. Which means you end up with an arms race of
I’ve been reading the sandboxing arguments between infosec people and AI alignment folks, and I tried to summarize and referee them a bit in this post. blog.cryptographyengineering.com/2026/09/30/is-…
