Uber Monitors Its AI Agents Like Employees. Your Rollout Should Too

Uber is monitoring its AI agents the way it monitors people. Chief Information Security Officer Philip Martin argues that insider-threat practice built for employees transfers to models, and that a program for AI will be essential to deploying them safely. Uber handles more than 40 million trips and delivery orders a day, so this is not a thought experiment.

The mechanics are specific. Uber open-sourced AI Detection and Response (ADR), which collects real-time telemetry from deployed agent harnesses and can act on it. An internal LLM gateway sets the perimeter and governs usage. Martin names two priorities for defenders: real-time detection and intervention, such as blocking a risky tool call, and discovering emerging model behaviors that need new safeguards. Anthropic’s safeguard outputs will feed into ADR.

The telemetry layer is becoming standard. Anthropic just put OpenTelemetry trace export from Claude Managed Agents, the Messages API and the Claude SDKs into public beta, covering every agent turn, model call, tool use and MCP server call. FloQast places each model call in the same Grafana Cloud trace as its accounting workflow, so it can answer auditors about what an agent did.

That sets the order of operations for a rollout: instrument first, then decide how much access an agent gets. Companies that grant access first and ask what happened later have nothing to show an auditor. If you are going to wield agents inside a business, three questions come before any permission: where does the telemetry land, who reviews it, and what can block a call in real time?

Uber’s answer is a gateway plus ADR. Yours can be smaller — a single trace destination and a named reviewer is a start — but it needs an answer before the first agent ships. Outcomes depend on being able to see what the agent did.