Azure OpenAI: Recurring HTTP 400 Failures During Authorized Cybersecurity Analysis

Problem description

I am experiencing recurring failures with Azure OpenAI during authorized defensive cybersecurity analysis. Some requests return HTTP 400 errors labeled as 'invalid_request' or 'cyber_policy', while others succeed. The failures seem prompt-dependent, with benign prompts working and certain security analysis prompts failing. I am seeking assistance to understand the cause and resolve this issue.

Environment

Microsoft Foundry Azure OpenAI resource used for authorized defensive cybersecurity analysis, tested from both a local environment and an Azure virtual machine.

What I've already tried

I identified the start of the issue as September 1, 2026. I reproduced the behavior from both local and Azure VM environments, comparing prompt types. Benign prompts succeed, while certain cybersecurity prompts fail. I reviewed response headers during failures, noting no HTTP 429 throttling errors and that capacity appeared available. I confirmed that the account has approvals for Modified Guardrails and Modified Abuse Monitoring, and there is a separate approval on the OpenAI platform for Trusted Access for Cyber under Daybreak Blue.

Current status

The issue persists, with some cybersecurity analysis requests failing and benign requests succeeding. I am seeking guidance to determine if a cyber policy block is causing this behavior and to understand any additional approvals or configurations needed to resolve the issue.