Researchers report agent defense that reduces harm without increasing missed emergencies

Researchers report that a trust-gating defense produced the lowest measured harm on every model tested in SentryBench, a new benchmark for AI agents using public-safety tools. The preprint evaluates attacks alongside failures to respond to genuine emergencies and unwarranted actions. Across 840 scored episodes involving three frontier models and three defenses, encouraging greater caution drove missed emergencies to 100 percent on two models for at most a 3.7-point change in attack success. The trust-gating wrapper kept missed emergencies at or below baseline while improving containment. The authors also found that provider content filters performed part of the defense, requiring separate reporting.