Azure OpenAI Abuse Monitoring

Prateek Dorwal 0 Reputation points
2025-06-24T06:49:54.8266667+00:00

Does Azure OpenAI monitor prompts for Abuse or other potential harmful content for Unmanaged customers? I know it does for Managed customers, but I could not find any clear answer if it does for Unmanaged customers as well?

Azure OpenAI in Foundry Models
0 comments No comments

2 answers

Sort by: Most helpful
  1. Abiola Akinbade 30,500 Reputation points Volunteer Moderator
    2025-06-24T08:06:09.5966667+00:00

    Hello Prateek Dorwal,

    Thanks for your question

    Simple answer: Yes it does

    This monitoring is active for all customers unless they have been approved for modified abuse monitoring. Unless you are a managed customer with approved modified abuse monitoring, Azure OpenAI’s abuse monitoring system will be active for your account

    See:

    https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/abuse-monitoring

    For modifications you need approval with this form
    https://customervoice.microsoft.com/Pages/ResponsePage.aspx?id=v4j5cvGGr0GRqy180BHbR7en2Ais5pxKtso_Pz4b1_xUOE9MUTFMUlpBNk5IQlZWWkcyUEpWWEhGOCQlQCN0PWcu

    You can mark it 'Accept Answer' and 'Upvote' if this helped you

    Regards,

    Abiola

    Was this answer helpful?

    0 comments No comments

  2. Pavankumar Purilla 11,655 Reputation points Moderator
    2025-06-24T08:01:50.8933333+00:00

    Hi Prateek Dorwal,

    Yes, Azure OpenAI does monitor prompts and completions for abuse and potentially harmful content even for unmanaged customers. The abuse monitoring system is applied by default across all Azure OpenAI customers whether managed or unmanaged. This includes real-time classification of both input prompts and model outputs to detect harmful categories such as hate speech, violence, sexual content, or attempts to jailbreak the model. Azure OpenAI retains flagged prompts and completions for around 30 days to enable pattern analysis, automated and human reviews, and enforcement actions if needed. If a customer requires stricter confidentiality for example, to ensure that prompts and completions are not stored or reviewed by humans they can apply for modified abuse monitoring, which disables storage and human review but still keeps automated abuse detection active. This ensures that even unmanaged customers are covered by Microsoft's responsible AI safeguards, unless they have received approval for modified monitoring through a formal process.

    Was this answer helpful?


Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.