Uber outlines a three-layer safety design for AI agents

Uber Engineering describes a Secure AI design that checks whether an AI agent's actions match what the user asked for, not only whether the agent has permission to act.

According to a repost by Meng Shao, the design has three layers: pre-launch review through AI Registry and Argus, runtime action checks through AI Guard, and continuous testing using replay and red teaming.

Uber has not published block rates, false-positive rates, or latency data, so the post describes deployment experience rather than measured results.