Nordic guardrails: Azure PII and Content Safety on the Gravitee AI gateway
Sam
Summary
I recorded this video to show how to bypass current limitations in Azure's native guardrails by combining Azure PAI filtering, Content Safety, and Azure Language services with a custom LLM proxy. I set up three UK‑South resources—Content Safety for harmful content and prompt injection, Language for personal data detection, and Azure OpenAI for the downstream model—then wired them together in Policy Studio using two Groovy scripts that shape requests, call the Azure endpoints, and enforce interrupt policies that return 403 when violations are detected. The demo runs sample traffic, showing organization masking, phone‑number redaction, and blocked Norwegian threats, with logs confirming 35/39 harmful prompts blocked and low latency (~270 ms). I note gaps in Finnish support and deep multi‑turn injection detection, and invite follow‑up discussions.