Source: Foundry Models sold by Azure — abuse monitoring

Microsoft's description of abuse monitoring for Azure OpenAI and other models sold by Azure: classification, pattern capture, LLM-then-human review, and the Limited Access application for modified abuse monitoring.

Contents

Summary

Explains the four stages of Azure’s abuse-monitoring pipeline and how customers with sensitive workloads can apply to turn off its storage and human-review parts.

Key claims

  • Stages: content classification, abuse-pattern capture (frequency, severity, intent), review and decision, then notification and action by email.
  • Flagged prompts and completions are reviewed by automated means (including LLMs) by default; that automated review does not store them. Human review happens only when automated confidence is insufficient or unavailable, via Secure Access Workstations and just-in-time approval; for EEA deployments reviewers are in the EEA.
  • Customers meeting extra Limited Access criteria can apply by form for modified abuse monitoring; some advanced models have stricter criteria.
  • With modified abuse monitoring, human review is not performed; Microsoft warns that abuse detection may be less accurate and customers must still respond to notifications.
  • The page gives no retention period for the human-review store.

Feeds

Data retention and residency

Suggest changes →