Uber CISO Argues AI Agents Need Insider Threat Programs, Leveraging Safeguards and ADR

Philip Martin, Uber's Chief Information Security Officer, contends that AI agents pose risks similar to human employees and therefore require equivalent insider threat programs. As frontier models like Claude Fable gain the ability to process long-term context and generate specific outputs, the potential for accidental misuse or actions outside expected constraints increases. Martin emphasizes that while AI adoption is a business imperative, security measures must evolve as models take on more complex, autonomous work.

Integrating Safeguards and Monitoring

Uber employs Anthropic's Enterprise Frontier Safeguards to balance model capability with data control. This framework allows Uber to protect customer data and trade secrets without granting Anthropic access to proprietary information, while giving Uber control over risk mitigation actions when safeguards are triggered. This approach complements Uber's recently open-sourced AI Detection and Response (ADR) tool, which gathers real-time telemetry from agent harnesses to detect when agents act outside their bounds.

Real-Time Detection and Future Architecture

The security strategy prioritizes real-time detection and intervention, aiming to warn users or block actions before harmful outcomes occur. By combining Uber's own telemetry from ADR with Anthropic's classifier outputs, the team creates a comprehensive detection environment similar to traditional cybersecurity setups. Martin envisions a future where teams of agents, each associated with a specific service, communicate threat models to one another, further enhancing security efficacy.