Uber CISO Urges Real-Time Controls for Frontier AI Systems

Uber Chief Information Security Officer Philip Martin said frontier AI systems need real-time monitoring and intervention similar to controls used to detect employee insider threats as models become more capable.

Martin said advanced models can absorb large amounts of context over long periods and produce more specific outputs. He said they may also act beyond expected constraints, making prompt- and tool-level oversight important for deployment.

Uber uses AI in security reviews and threat modeling by providing models with code, design documents, Jira issues and historical service information. Martin also described a potential system in which each service has an associated agent that exchanges threat-model information with other agents, mirroring how production services interact.

Uber’s AI security setup combines an LLM gateway that establishes a perimeter for AI use with its open-source AI Detection and Response, or ADR, tooling, and Anthropic’s Enterprise Frontier Safeguards. The safeguards allow Uber to retain control of its data and determine how to respond when one is triggered.

ADR is deployed in production at Uber. The production system inventories AI applications, monitors agents, tests defensive measures and flags potential threats. Its prevention component is not included in the current open-source release.

Martin identified two priorities for defenders: intervening in real time as close as possible to a prompt or tool call, and identifying emerging behavior patterns so new findings can be incorporated into safeguards. He described the process as a cycle in which new model behaviors are detected, countermeasures are added and additional behaviors emerge under the new constraints.

The focus on model behavior follows earlier coverage of AI safety incidents, including attempts to bypass safeguards, escape sandboxes and evade monitoring in training, evaluation and research environments.