A major study led by researchers at the Massachusetts Institute of Technology (MIT) examined 30 widely used “agentic” AI systems — autonomous programs that can plan and execute tasks rather than just respond to prompts — and uncovered serious safety and governance gaps in how they’re built and deployed. These systems are increasingly being integrated into workflows where they can access databases, send emails, and perform multi-step tasks with minimal human supervision, yet most vendors provide little to no documentation about how they ensure safety or control for harmful behavior.
According to the research, a large majority of these agentic systems lack formal safety policies, don’t undergo independent third-party safety assessments, and offer limited or no mechanisms to monitor activity or stop execution once they are running. Twelve out of the thirty systems reviewed provide no meaningful usage monitoring at all, leaving organizations blind to how resources are consumed and what actions the agents undertake.
The report highlights that these shortcomings are particularly troubling because agentic AI systems are moving beyond simple chatbots — they can interact with enterprise platforms and make changes in real time — yet transparency around testing, risk evaluation, and operational safeguards is inconsistent across the ecosystem. Some systems do better — for example, a few implement traceability mechanisms like cryptographically signing browser requests — but many do not disclose even the basics of how they manage risks.
Researchers and industry observers argue that this lack of oversight and disclosure could lead to serious vulnerabilities as agentic AI becomes more widespread in business and consumer environments. They suggest that developers need to establish stronger safety standards, more transparent evaluation processes, and clear boundaries for what agents can and cannot do — or regulators may step in to enforce such controls.
