sandbox-probe
What each AI coding agent's sandbox lets through — measured, not assumed. A cell is leaked only when the unconfined baseline could do it too. blocked = the sandbox closed the door; n/a = nothing to prove.
What each AI coding agent's sandbox lets through — measured, not assumed. A cell is leaked only when the unconfined baseline could do it too. blocked = the sandbox closed the door; n/a = nothing to prove.
Exposure = number of leaked capability categories (0–8). Vertical markers are version releases.
Every capability that flipped state between two configurations, attributed to what moved.