Nordic guardrails: Azure PII and Content Safety on the Gravitee AI gateway
Azure AI Language and Content Safety can guard a Nordic-language chatbot from the gateway today, masking personal data and refusing threats and prompt injection before the model is called
Pre-release buildInternal only
Made for Posten Bring
Say while it plays
Gravitee's native guardrail was masking organisation names as personal data and letting non-English threats and prompt injections straight through, so this chains two Azure services, Content Safety and Language, onto the LLM proxy as a stopgap ahead of native non-English support.
Two Groovy steps do the work: one shapes each turn into the callouts for PII detection and harm and injection scoring, the second decides whether to mask or block, and a blocked call returns a 403 with the downstream model never called.
Testing blocked 35 of 39 harmful Nordic prompts with no false positives, and the setup fails closed, so if Azure is unreachable the prompt never reaches the model at all.
Don't claim
Don't claim reliable Finnish coverage: the recording says Finnish is where it falls short, because Azure itself says it wasn't really trained on that language.
Don't claim deep multi-turn prompt injection is solved: the recording says hiding an injection deep within a multi-turn conversation isn't perfect either.
Don't present this as native Gravitee guardrail functionality: the recording calls it a workaround built from Azure callouts to get live before native non-English support lands in 4.14.