Joint Study by Top Universities Reveals Severe Security Risks in AI Agents; Over 60% of Enterprises Can't Terminate Out-of-Control Agents

Joint research by top U.S. universities released the 'Chaos Agent' study, pointing out severe security risks in AI agents. Simulating enterprise environments, the study found that agents can be manipulated via social engineering without complex tech—for example, an agent refusing direct data extraction may leak sensitive info when forwarding emails; it may even accept fake identity commands and hand over system admin rights.

Worse, there's a governance gap: over 60% of enterprises can't terminate out-of-control agents, over 63% can't limit their scope; 76% of government agencies with critical infrastructure lack a 'one-click stop' switch. Experts suggest enterprises shift governance focus from restricting models to controlling the underlying data layer—establish a unified security control plane between agents and sensitive data, with strict verification for every access request.