News
Philip Martin says AI agents need real-time monitoring and controls similar to employee insider-threat programs, citing Uber’s ADR and safeguards.
Uber’s chief information security officer, Philip Martin, said frontier AI systems need monitoring controls similar to the insider-threat programs companies already run for employees. In an interview published by Anthropic on Oct. 13, 2026, Martin argued that increasingly capable models can be directed to act outside expected constraints, making real-time detection and intervention a core requirement for safe deployment.
Evidence and context
Martin said the biggest change in frontier models is their ability to absorb large amounts of context over long time horizons and produce more specific outputs. At Uber, he said that capability is being applied to security reviews and threat modeling by feeding models code, design documents, Jira issues and historical service context. He also described a future in which each service has an associated agent that can exchange threat-model information with other agents, mirroring how production services interact.
The interview also says Uber worked with Anthropic on Enterprise Frontier Safeguards and that those safeguards allow Uber to keep control over its data and decide how to respond if a safeguard is triggered. Martin said Uber’s AI security setup combines its LLM gateway, which establishes a perimeter for AI use, with its open-source AI Detection and Response, or ADR, tooling and Anthropic’s safeguards. Uber’s ADR repository separately states that the tool is deployed in production at Uber and is designed to discover AI tools, observe agent activity, benchmark defenses and detect threats; the repository also notes that its prevention component is not part of the current open-source release.
Martin said defenders should prioritize two areas: first, intervening in real time as close as possible to the prompt or tool call; second, identifying emerging behavioral patterns and feeding those findings back into safeguards. He framed the approach as a cycle in which new model behaviors are detected, countermeasures are added, and other behaviors then emerge under the new constraints.
The comments are notable because they come from a vendor-published interview rather than an independent audit of Uber’s controls. The source presents Martin’s views on AI risk and Uber’s security architecture, but it does not provide performance metrics, incident data or third-party verification of the safeguards’ effectiveness. SecurityWeek separately reported Martin’s appointment as Uber CISO on June 26, 2026, but that date should not be treated as a verified start date for his role.
Be the first to comment