Securing Ai Agents

Authorization alone cannot tell whether an agent action matches what the user asked. Uber's approach of tracing privileged actions back to the original request via prompt provenance is a concrete defense against indirect injection and goal hijacking.

• MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
• guardrails — The checks around a model that block bad inputs and outputs — filters, validators, and permission rules the model itself can't override.
• red teaming — Deliberately attacking your own AI system to find what makes it fail before someone else does.
• AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

• articleAi Code Cleanup
• articleBuilding Ubers Redlining Agent
• articleDesigning Mcp Gateway
• articleMichelangelo Machine Learning Platform

• blogAI Security Is an Engineering Problem — How to Solve It at Every Layer of the Agent StackSaša Zdjelar
• articleTrustworthy AgentsAnthropic research
• articleSolving The Agent Identity CrisisUber Engineering 2

Checking sign-in…

Loading comments…