5 Aug 2026

Uber has built a production security system specifically for AI agents, and the financial services industry should pay close attention. Not because Uber is a bank, but because they are months ahead of most lenders in understanding what happens when AI agents actually do things in the real world.

Most organisations deploying AI in consumer finance are still thinking about safety in terms of model behaviour: will it say something harmful, will it hallucinate a rate, will it discriminate. Those are real concerns. But ADR is solving a different problem entirely. When an AI agent can read data, call APIs, write code, and take actions autonomously, the threat model looks more like application security than content moderation.

Think about what we are already building. AI support bots that access customer account data. Coding assistants used by engineers who have access to loan origination systems. Agents that may soon have write permissions, not just read access. The attack surface is real and it is growing faster than most security teams have noticed.

The four-capability model Uber describes is worth understanding:

• Observe what agents are actually doing in production
• Evaluate whether defences are holding
• Detect anomalous or adversarial behaviour
• Block unsafe actions before they execute

This is standard security thinking applied to a new class of system. The fact that it needed to be rebuilt from scratch tells you how immature the tooling is right now.

For UK consumer credit specifically, the FCA's increasing focus on operational resilience and AI governance means this will eventually become a compliance question, not just an engineering one. The firms that treat agent security as an afterthought will be explaining incidents to regulators rather than preventing them.

Uber open-sourcing parts of this is genuinely useful. But the more important signal is that a company with serious engineering resources found this problem hard enough to publish a paper about it. If it is hard for them, what does that mean for a mid-sized UK lender trying to deploy agents with a fraction of that capability?