Securing AI Agents [at Uber]: Lessons from Real-World Deployment (www.uber.com)

Uber has announced a new defense-in-depth security framework for AI agents, addressing the unique challenges posed by agentic AI's autonomous reasoning capabilities. Traditional security models, which focus primarily on identity and authorization, are inadequate for governing the nuanced behaviors of AI agents that can dynamically interpret content and perform real-world actions. The developed framework, which encompasses provisioning-time validation, runtime enforcement, and continuous assurance, aims to ensure that actions taken by agents are not only authorized but also contextually appropriate and aligned with user intents. This approach mitigates risks such as prompt injection and behavioral drift by embedding strict semantic and behavioral controls within the agent architecture itself. Significantly, Uber's architecture integrates seamlessly with its existing AI development stack, allowing security measures to be inherited throughout the agent lifecycle—from creation to deployment. By utilizing specialized local models for quick decision-making and continuous calibration based on agent behavior and risk profiles, Uber addresses the performance and latency challenges typically associated with traditional systems. The framework also supports comprehensive observability and auditing, enabling effective threat modeling and red teaming to adapt safeguards as both agents and attack techniques evolve. This proactive strategy not only enhances agent safety but also provides practical insights for the AI/ML community in designing robust security systems for emerging AI technologies.