Checking the cage: a run-time seal model for AI evaluation sandboxes
Published Sep 2026The first paper out of Nullmoor: the catalogue of 48 escape routes a sandbox has to close, the rule that every probe must be able to fail, and the five false seals we found by attacking our own checker. Coverage is stated as a fraction — 36 of 48 routes fully watched — with the uncovered remainder named rather than rounded away. A paper that reports only successes is not a paper.