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

A comprehensive study led by MIT and University of Cambridge researchers has exposed significant security and transparency gaps in today's agentic AI systems. The 39-page report, 'The 2025 AI Index: Documenting Sociotechnical Features of Deployed Agentic AI Systems,' analyzed 30 major AI agent platforms including OpenAI's ecosystem, Perplexity's Computer, IBM's watsonx, and Alibaba's MobileAgent. The findings reveal a troubling pattern: most systems disclose nothing about safety testing protocols, third-party evaluations, or potential risks. This lack of transparency comes as agentic technology moves mainstream, highlighted by OpenAI's recent hiring of Peter Steinberg, creator of the powerful but flawed OpenClaw framework that can hijack personal computers.

The technical analysis shows specific vulnerabilities: 12 out of 30 agents provide no usage monitoring or only notify users at rate limits, making resource tracking impossible for enterprises. More alarmingly, systems like Alibaba's MobileAgent, HubSpot's Breeze, IBM's watsonx, and n8n's automations lack documented stop options despite autonomous execution capabilities. Most agents also fail to disclose their AI nature to end users or third parties by default, bypassing protocols like watermarking or robots.txt identification. The researchers identified persistent limitations across eight disclosure categories, with most systems offering no information about potential risks, monitoring capabilities, or execution trace tracking. This creates what the study calls a 'security nightmare' where rogue agents could operate unchecked.