Classification:

attack

Tactic:

TA0005-defense-evasion

Technique:

T1562-impair-defenses

VIEW RULE IN DATADOG VIEW SIGNALS

Goal

Detects modification of an Azure AI Foundry Responsible AI (RAI) content-filter policy.

Strategy

This rule monitors Azure Activity Log events where @evt.name is MICROSOFT.COGNITIVESERVICES/ACCOUNTS/RAIPOLICIES/WRITE and @evt.outcome is Success. This may indicate that LLM guardrails such as content filters and prompt injection protections were weakened for an LLM deployment.

Triage & Response

• Examine the resourceId and properties.entity fields on the triggering event to identify which Cognitive Services account and RAI policy {{@usr.id}} modified.
• Review the specific policy change by comparing the modified policy’s content-filter categories and severity thresholds against the account’s baseline configuration.
• Verify if {{@usr.id}} has a legitimate business reason to modify RAI policies for this account, such as a documented content-filter tuning request.
• Check whether the modification loosened restrictions, such as raising severity thresholds or removing blocked categories.
• Identify any subsequent inference requests against the affected deployments and account shortly after the policy change to assess whether the modification was followed by suspicious model usage.