Uber Enhances AI Agent Security with New Identity Platform

Uber's recent efforts to enhance the security of AI agents reveal a strong commitment to tackling the complexities of identity management in autonomous systems. Traditional models have struggled to effectively address the dynamic nature of AI agents, leading the company to outline a new internal platform designed to resolve what it calls the "identity crisis" for these technologies.

The Challenge of Identity for AI Agents

The core issue stems from conventional identity models that cater to static workloads and human users. As AI agents increasingly serve as intermediaries executing tasks that involve multiple actions and workflows, the absence of clear attribution becomes problematic. Uber has identified a phenomenon called the "Agency Gap," where downstream systems recognize only generic service identities, obscuring the true initiator of actions. For example, if an on-call engineer uses an agent to resolve a system issue, the resulting pull request may lose the engineer's identity, complicating audits and compliance efforts.

The Zero Trust Foundation Approach

To address these challenges, Uber has expanded its existing Zero Trust Architecture. This approach focuses on establishing a verifiable cryptographic identity for AI agents, ensuring that authorization is enforced at every step of the agent's interactions with downstream systems. Key components of this architecture include an Agent Registry, an AI Agent Mesh, and a Security Token Service (STS) that issues dynamic, short-lived tokens to agents as they operate.

The STS acts as a trust broker, providing JSON Web Tokens (JWTs) that are scoped for specific destinations and contain an embedded history of the agent's actions. This setup allows for end-to-end traceability from the original user to the final action performed by the agent, addressing concerns about the provenance of actions in complex workflows.

Implementing Agent Identity in Action

Uber's new platform also integrates with the Model Context Protocol (MCP) Gateway, which serves as a policy enforcement point for agent interactions. This gateway verifies agent identities and ensures that appropriate policies are followed during tool invocations. With the addition of AI Guard—designed to detect prompt injections and redact personally identifiable information—Uber is significantly enhancing the security of its AI systems.

When an engineer interacts with an Oncall Agent, their identity is preserved as the request is initiated. The Oncall Agent then connects with the STS to obtain a new JWT, which is contextualized for the next agent in the workflow, such as an Investigation Agent. The resulting token carries a verifiable lineage of actions, ensuring that downstream systems are aware of the entire context of the request.

Future Directions

As Uber advances these initiatives, the company is also monitoring emerging industry standards, including those from the IETF WIMSE working group. This proactive stance ensures that Uber’s developments remain in line with broader industry trends in AI and security.

To support a consistent implementation of its security measures, Uber has developed a Standardized Agent-to-Agent (A2A) Client. This tool automates the exchange of STS JWTs and actor chain propagation, fostering a secure and efficient environment for developers.

Uber's new platform marks a significant step forward in securing AI agents by addressing identity and accountability challenges. As the field of autonomous agents evolves, these measures will be essential for maintaining trust and compliance in AI-driven processes.