Uber checks if each AI agent step fits the task of the agent
Uber adds a check that tells if an agent step is right in context, and not only if the agent has permission.
Uber Engineering shows a framework that keeps AI agents safe when they read external content and do tasks in the world. Identity systems can tell if an agent can use a resource. They cannot tell if one step is right in context. Uber uses three layers: registration before launch, checks at runtime and inspection after launch. Uber gave no test numbers.
How it works
Each agent goes in the Uber AI Registry with its owner, its permitted intents and its limits. A tool, Argus, examines the code and the configuration of the agent for weak points. At runtime, guards check each step against policy. They give a verdict before the step continues. After launch, a record links steps, tool calls and protection decisions. Teams can replay recorded work when the rules change.
The proof
Uber gave no test numbers. Uber tells that clear limits for each agent give good results. General limits did not work, because each agent has a different purpose and risk.
What is not known
Tristan G., a technology officer at the UN, tells that the work grows with the number of workflows. The officer wants to know how this works for vendor agents, where you cannot add your hooks. Uber also lists incorrect alerts, latency and developer work as costs.
Insight: Identity checks tell who an agent is, not if its step is right
The Chain of Trust, from GoDaddy, Infoblox, Google and Cloudflare, checks who an agent is. Uber tells that this is not sufficient. Tests of AI browser agents found that bot defenses from before AI agents miss them. This reads as a divide: identity checks go open, but checks of agent behavior stay in each company. Control goes to the company that runs the agent.
Sources
Posted