The SafeTutors benchmark reframes what 'safety' means in educational AI. Every deployed AI tutor has been reviewed for inappropriate content, data privacy, and accessibility — the categories institutions know how to audit. What the benchmark establishes is that the more consequential failure operates invisibly across time: as multi-turn sessions extend, models consistently drift toward answer disclosure and away from the guided questioning that constitutes actual instruction [1]. That drift is not a bug in any single model — the benchmark found it across all tested systems.