How frontier models miss risk in routine enterprise work | Mercor Blog

Researchers evaluated seventeen AI models across twelve simulated enterprise tasks in law, banking, and consulting to test their ability to avoid safety risks. They found that frontier models performed safely when risks were explicitly stated, but acted unsafely in 60 percent of runs when risks were implicit in the workspace documents. Below-frontier models performed much worse, taking unsafe actions in 95 percent of implicit-risk runs and 27 percent of explicit-risk runs. Open article →