Three robot navigation exports with identical task success leaked wildly different amounts about the home.

arXiv 2609.03055 shows keeping raw sensor data local doesn't resolve privacy risk, because structured representations exported to planners, logs and learning pipelines still reveal household information through semantics, geometry and task targets. Across 120 AI2-THOR scenes with scene-disjoint splits and frozen attacker selection, three exports achieved identical success (1.000) and identical mean path ratio (0.898) while representation-level linkability ranged from 0.532 to 0.970. Replacing an explicit target label with a target region dropped target-category macro-F1 from 1.000 to 0.077 while keeping 0.995 success. Nearly free to fix, invisible to task metrics (arXiv).