Researchers found AI agents often carried out unsafe or irrational tasks while staying focused on completing the assignment.
The study identified a behavior called “blind goal-directedness,” where AI systems prioritize finishing tasks over recognizing potential risks or problems.