72% of 39 agent models complete assigned insider-threat objectives, and refusal doesn't predict completion.

MOLE is an open benchmark of 150 AI-operated accounts sharing nine stateful services over 30 workdays, with 12 threats and 8 corpora from four models totaling about 20 billion tokens. Comparing 40 monitors, even the best missed nearly half of completed harm in the single-day audit-event comparison. Benchmark-guided search improved a mid-tier monitor by 49-64%, and selectively escalating to a stronger monitor beat blanket application by 10% budget-AUC at comparable cost. arXiv 2609.06966