Content Filter Severity Downgrade - Previously Blocked Inputs Now Passing with Lower Severity

Question

Hello Azure Support Team,

We are experiencing an unexpected change in the Azure OpenAI Content Filter severity levels for our deployment.

Issue Summary: Certain user inputs that were previously blocked with "high" severity violations are now being allowed through with only "low" severity warnings. We have not made any changes to our content filter configuration, but the filter appears to have become less strict.

Details:

• Deployment Name: lg-openai-gpt-4.1-mini

Model: GPT-4.1-mini

• LOCAL_OPENAI_API_VERSION = "2024-05-01-preview"

LOCAL_AZURE_OPENAI_ENDPOINT = "https://lg-openai-ins.openai.azure.com/"

Example Inputs:

"IT 관련 최신 뉴스 나온거 있어" (Are there any latest IT-related news?)

"인터넷에서 예맨사태 관련 뉴스 확인해줘" (Check news about Yemen situation on the internet)

Observed Behavior Change:

Previously: These inputs were blocked with "high" severity (violence category)

Currently: These inputs pass through with only "low" severity warnings

Current Response:

{

Copyjson

Questions:

Has there been a recent change to the content filter model or severity classification algorithm?

Why has the severity level been downgraded from "high" to "low" for the same inputs?

Is this an intentional policy change or a potential issue with the content filter?

How can we ensure consistent content filtering behavior for our production service?

Is there a way to review content filter configuration change history or logs?

Concern: We need to understand whether this is:

An improvement in the content filter accuracy (correctly identifying these as low-risk)

A degradation in filter sensitivity (potentially allowing harmful content)

A temporary issue that may revert

Impact: This inconsistency affects our ability to predict and handle content filtering in production. We need to understand the root cause to properly configure our error handling and user experience.

Could you please investigate this change and provide clarification on the content filter behavior?

Thank you for your assistance.

Best regards, [Your Name] [Your Company]Hello Azure Support Team,

We are experiencing an unexpected change in the Azure OpenAI Content Filter severity levels for our deployment.

Issue Summary:
Certain user inputs that were previously blocked with "high" severity violations are now being allowed through with only "low" severity warnings. We have not made any changes to our content filter configuration, but the filter appears to have become less strict.

Details:

Deployment Name: lg-openai-gpt-4.1-mini

Model: GPT-4.1-mini

• LOCAL_OPENAI_API_VERSION = "2024-05-01-preview"

LOCAL_AZURE_OPENAI_ENDPOINT = "https://lg-openai-ins.openai.azure.com/"

Example Inputs:

"IT 관련 최신 뉴스 나온거 있어" (Are there any latest IT-related news?)

"인터넷에서 예맨사태 관련 뉴스 확인해줘" (Check news about Yemen situation on the internet)

Observed Behavior Change:

Previously: These inputs were blocked with "high" severity (violence category)

Currently: These inputs pass through with only "low" severity warnings

Current Response:

{

Copyjson

Questions:

Has there been a recent change to the content filter model or severity classification algorithm?

Why has the severity level been downgraded from "high" to "low" for the same inputs?

Is this an intentional policy change or a potential issue with the content filter?

How can we ensure consistent content filtering behavior for our production service?

Is there a way to review content filter configuration change history or logs?

Concern:
We need to understand whether this is:

An improvement in the content filter accuracy (correctly identifying these as low-risk)

A degradation in filter sensitivity (potentially allowing harmful content)

A temporary issue that may revert

Impact:
This inconsistency affects our ability to predict and handle content filtering in production. We need to understand the root cause to properly configure our error handling and user experience.

Could you please investigate this change and provide clarification on the content filter behavior?

Thank you for your assistance.

Best regards,
jaehyung1.lee

lge electronics

Answer 1

Hi 이재형/선임연구원/AI서비스개발팀

Thank you for confirming that issue was resolved with provided content filter setting .

Attached case summary for reference.

Issue

Customer suspected that sensitive prompts were passing with high threshold content filter

Recommendation shared

1. Took the concerned prompts
2. and Shared input and output content filter setting that blocked sensitive prompts as per design.
3. Tested Langchain Azure OpenAI code with content filter enabled on deployment and shared confirmation on support ticket.

Reference used - https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/content-filters?view=foundry-classic

Status - Customer confirmed case resolution with additional queries on Default and DefaultV2 content filter addressed

Please take a minute to accept this answer if you appreciated our inputs.

Thank you for using Azure Services.