Is Azure Content Saftety Filter enabled by default for the LLM models /AI servicesdeployed through the platform ?

Question

Is Azure Content Saftety Filter enabled by default for the LLM models /AI servicesdeployed through the platform ?

Premnath S 20

Wanted to understand if Azure content safety filtering is enabled to monitor on the prompts and responses to LLMs/AI services hosted through the platform ? Seems confusing for me to understand from the Azure learning articles on this as somewhere it seems to have been mentioned as to be enabled by default and other places, it is mentioned to be limited to only few categories(Hate and Fairness, Sexual, Violence and Self-harm) to be enabled by default.

So wanted to primarily understand if the other 3 categories(Prompt Shields, Protected material, Groundedness) are enabled by default?

Thanks

Accepted answer

0 additional answers

Your answer

Answer 1

SriLakshmi C 5,205 Microsoft External Staff Moderator

Hello @Premnath S,

Azure Content Safety Filter is enabled by default for Large Language Models (LLMs) and AI services deployed through the Azure platform, particularly when using serverless APIs. This default configuration applies to four main harm categories: Hate and Fairness, Self-Harm, Sexual Content, and Violence. These categories are enforced automatically to ensure a baseline level of safety and content moderation for end users.

However, other advanced categories such as Prompt Shields, Protected Material, and Groundedness are not included in the default configuration.

To manage these settings more granularly, Azure offers configuration controls at the resource level, such as through Azure AI Studio or AI Foundry. This allows teams to customize which categories of content filtering are active, depending on the sensitivity of the application or target audience.

For the most current information and capabilities, it's recommended to regularly check the official Azure documentation and "What’s New" updates in the Azure AI portal, as the platform continues to evolve and expand its safety feature set.

Please refer this Default content safety policies for Azure AI Model Inference.

Default content safety policies.

I Hope this helps. Do let me know if you have any further queries.

If this answers your query, please do click Accept Answer and Yes for was this answer helpful.

Thank you!

Premnath S 20 Reputation points

2025-04-24T10:45:56.65+00:00

Thanks for clarifying , had asked cause I also had a similar understanding with harm categories being enabled by default but then cam across this page wherein it mentions some of the prompt shield and protected material categories to be enabled by default.
https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/content-filters
SriLakshmi C 5,205 Reputation points Microsoft External Staff Moderator

2025-04-24T11:56:57.6633333+00:00

Hi Premnath S,

You're absolutely right to point out that Azure's default content safety filters can seem a bit confusing, especially when comparing the traditional harm categories with newer safety layers like Prompt Shields and Protected Material.

To clarify, Azure Content Safety is enabled by default for all Azure OpenAI deployments particularly for serverless endpoints. This default configuration includes four core harm categories: Hate and Fairness, Self-Harm, Sexual Content, and Violence. These are consistently enforced unless explicitly modified based on subscription level or access permissions.

In addition to these, other safety features such as Prompt Shields (to detect prompt injection or jailbreak attempts), Protected Material (such as PII or intellectual property), and Groundedness (to assess factual alignment of the response with source input) are partially enabled depending on the model version and deployment type. As noted in the documentation you referenced, some sub-features—like basic prompt safety checks and PII detection are active by default in newer deployments.

However, full enforcement or fine-tuning of these additional layers often requires manual configuration at the resource level using Azure AI Studio or Azure AI Foundry. This allows for more granular control based on your application’s safety requirements.

I Hope you understand.

If this answers your query, please do click Accept Answer and Yes for was this answer helpful.

Thank you!

Share via

Is Azure Content Saftety Filter enabled by default for the LLM models /AI servicesdeployed through the platform ?

0 additional answers

Your answer