Microsoft Foundary - guardrails

Hi Tony Healy ,

Thank you for the question. I understand that organisations working in areas such as coronial investigations, public safety, healthcare, or law enforcement may legitimately process content that includes violence, death, or drug-related topics.

Microsoft does not typically lower or disable guardrails on a subscription-by-subscription request through Azure Support alone. However, Azure AI Foundry content filters are configurable, and certain advanced filtering configurations require approval through Microsoft's Limited Access review process. Customers can adjust supported severity thresholds and, for eligible scenarios, apply for modified content filtering capabilities.

If your business workload is generating false positives despite using the least restrictive supported settings, we recommend:

1. Collecting examples of legitimate prompts/responses that are being blocked.
2. Reviewing your current content filter configuration and thresholds.

Please note that approval is subject to Microsoft's Responsible AI requirements and programme eligibility, and is not automatically granted for all subscription

Hello @Tony Healy

Yes, Microsoft Foundry Guardrails can be configured to be less restrictive, but there are limits to what you can change without additional Microsoft approval.

For the standard harm categories: Violence, Hate, Sexual, and Self-harm, Foundry allows you to configure severity thresholds independently for prompts and model responses. The default configuration generally blocks content classified as medium or high severity.

For your use case, where references to death, violence, or drug use are legitimate parts of professional case material, you could first configure the relevant categories so that only High severity content is blocked. That lets Low and Medium classifications through while still protecting against the highest-severity content.

In Foundry, review the Guardrails/content filter associated with the model deployment and adjust the input and output thresholds individually. I would also inspect the content_filter_results returned by rejected requests to determine exactly which category and severity are causing the rejection before changing the policy. Microsoft exposes the category, severity (safe, low, medium, high), and whether filtering occurred in the API results.

If High-only filtering still blocks legitimate material, Microsoft does provide a process for requesting Modified Guardrails. An approved configuration can allow options such as No filters or Annotate only, where the content is classified but isn't automatically blocked.

However, there is an important eligibility restriction: Microsoft currently states that Modified Guardrails are available only to customers/partners managed by a Microsoft account team or participating in an eligible program, and approval is subject to additional requirements. It isn't simply a subscription-level switch that Azure Support can disable on request.

For this particular scenario, I'd therefore recommend:

Capture several legitimate requests that are currently rejected, including their filter category/severity.

Configure the affected categories to the least restrictive standard setting appropriate for the workload (for example, High-only blocking).

If legitimate coroner-related material is still rejected, contact your Microsoft account team and request the Modified Guardrails/content filtering Limited Access review, explaining the regulated/professional nature of the workload and the controls around access to the data.

If approved, consider Annotate only rather than disabling filtering completely. Your application can then apply its own workflow-specific controls based on the returned classifications.

One other distinction is worth mentioning: Modified Guardrails and modified abuse monitoring are separate capabilities. Changing content-filter thresholds doesn't automatically disable Microsoft's abuse-monitoring mechanisms. Microsoft has a separate approval process for customers that need modified abuse monitoring.

Sharing these references with you:

Microsoft - Configure content filters in Foundry

Microsoft - Limited access and Modified Guardrails

Microsoft - Guardrails and controls overview

So, yes, Microsoft can approve less restrictive Guardrails, but I'd first determine whether High-only filtering solves the problem. If you need to disable filtering or change it to Annotate-only, that requires Microsoft's Modified Guardrails approval process.

Help make this community better for everyone: if this answer resolved your issue, please accept it or leave an upvote. If not, share more details in a comment so we can continue the discussion and find the right solution.