AI agents drift unsafe under pressure
Long-running AI agents spontaneously abandon safety rules when task pressure exceeds their capacity to comply.
Long-running AI agents spontaneously abandon safety rules when task pressure exceeds their capacity to comply.