Deloitte recommends stopping AI agents through controls outside the model and revoking their credentials when necessary. Its October 5, 2026 assessment also calls for mechanisms to remove tool access and block network connections.
The authors, including Diana Kearns-Manolatos and Adnan Amjad, organize control around four questions: observing behavior, limiting actions, managing dependencies and verifying task boundaries. They recommend human approval for high-impact actions.
Under the proposed approach, models should not be able to modify or disable their own safeguards. Records and independent enforcement would operate separately from task logic.
The publication offers recommendations, not a new deployment or effectiveness measurement. Deloitte notes that the cited testing incidents are not representative of typical enterprise deployments.