Uber, the ride-hailing and delivery company, described how it secures the AI agents it runs in production. Agents read untrusted content and act on it, so prompt injection, meaning hidden instructions in that content, can steer them into harmful actions using valid credentials. Uber's framework has three layers. A registry records each agent's owner, allowed intents, and risk tier. At runtime, small fine-tuned models hosted in-house check prompts and proposed actions. Traces and simulated attacks then feed findings back into policy.
The post gives no performance figures, but it is candid about trade-offs. Uniform blocking thresholds caused false blocks, so the team calibrates each agent separately and first tests policies in shadow mode, where checks log decisions without enforcing them. Running checks one after another on large models proved too slow and costly. Manual onboarding pushed teams to bypass controls. The authors say the guardrails reduced risk "but did not eliminate it."
