Responsible AI Content Filter Policy
Does ZopNight manage Responsible AI Content Filter Policy?
Responsible AI content-filter policies define which content categories an Azure OpenAI deployment filters and at what severity threshold. Policies are free, since no meter attaches, but ZopNight inventories each one with its mode (Blocking, Deferred, Default) and type, then links it to the deployments it governs for AI governance review.
Rules that fire on Responsible AI Content Filter Policy
No active rule family targets Responsible AI Content Filter Policy today. Rules that used to are retired, and retired rules publish no pages and fire no findings. Scheduling and permissions coverage are unaffected.
At a glance
| Field | Value |
|---|---|
| Scheduling notes | discovery only. |
Responsible AI policies define the content-filtering configuration applied to Azure OpenAI deployments. They are free but essential governance context for AI estates.
Content filter policies carry no meter
A RAI policy is an ARM child resource of a Cognitive Services account (raiPolicies under the account) and it never generates a charge. What it carries is posture: which content categories the filter screens, at what severity thresholds, and whether the policy blocks, defers, or merely annotates. Two estates with identical model deployments and identical bills can have completely different risk profiles depending on nothing but these objects, which is why they belong in the same inventory as the deployments themselves.
Policy mode and type in the inventory
Discovered via the AI enricher and linked to the deployments they govern, supporting AI governance visibility. Per policy, ZopNight records the mode (Blocking, Deferred, or Default) and the policy type, which separates Microsoft’s built-in policies from UserManaged ones a team authored. RAI sub-resources exist on any Cognitive Services account kind, so the sweep covers OpenAI and AIServices accounts alike, and the policies are grouped under a synthetic Content Filter Policies node per account to keep large estates readable.
Filter gaps an inventory exposes
The findings that matter here are governance findings. Custom policies that loosened default thresholds during an experiment and were never re-tightened before production traffic arrived. Deployments attached to a permissive UserManaged policy while the team believes the Microsoft default applies. And policy sprawl: dozens of near-identical custom policies across accounts, none documented, each a separate answer to an auditor’s question about what filtering was in force on a given date. An inventory with mode and type per policy turns each from a discovery-during-incident into a review item.
Content filters in Azure AI Foundry
The AI Foundry portal edits these under Safety + security: content filters are created there and attached to model deployments. From the ARM side they enumerate as raiPolicies children of the Cognitive Services account. That is the path automated discovery uses, and the one that catches policies the portal view never surfaces.