AI agents are fast, loose, and out of control, MIT study finds

A ZDNET article reviews a multi-institution MIT-led survey of 30 agentic AI systems, finding pervasive security and governance shortcomings, including minimal risk disclosure, weak monitoring, poor transparency about evaluations, lack of clear stop controls, and inconsistent identification of AI to third parties. The piece discusses specific products (e.g., OpenAI Atlas Agent, Perplexity Comet, IBM watsonx Orchestrate, HubSpot Breeze) and includes company rebuttals, arguing that developers must assume responsibility for documenting, auditing, and governing agent behavior as capabilities and risks grow.