MIT Study Finds AI Agents Lack Adequate Safety Controls and Oversight

Summary

A new MIT study highlights significant concerns around the safety and controllability of AI agents — AI programs that go beyond single-prompt interactions by connecting to external resources such as databases, email systems, and inventory platforms, and operating with a degree of autonomy to complete multi-step workflows. The research examines how these agents are being deployed in enterprise settings for tasks like processing purchase orders, managing inventory lookups, and automating customer