---
title: Azure AI Foundry anomalous inference volume via anonymous key
description: Datadog, the leading service for cloud-scale monitoring.
breadcrumbs: >-
  Docs > Datadog Security > OOTB Rules > Azure AI Foundry anomalous inference
  volume via anonymous key
---

> For the complete documentation index, see [llms.txt](https://docs.datadoghq.com/llms.txt).

# Azure AI Foundry anomalous inference volume via anonymous key
Classification:attackTactic:[TA0040-impact](https://attack.mitre.org/tactics/TA0040)Technique:[T1496-resource-hijacking](https://attack.mitre.org/techniques/T1496) 
## Goal{% #goal %}

Detects an anomalous spike in inference-endpoint call volume against an Azure AI Foundry (Cognitive Services) model deployment where the caller authenticated with an anonymous Azure OpenAI key. Anonymous keys are explicitly discouraged by Microsoft and are often used in LLMJacking attacks due to their lack of fine-grained permission scopes or auditability.

## Strategy{% #strategy %}

This rule monitors Azure Cognitive Services diagnostic logs for model inference calls where `@evt.name` is `ChatCompletions_Create`, `chatcompletions_create`, `create-response`, or `Create-Response`, restricted to events where `@properties` contains an empty `callerObjectId` (`"callerObjectId":""`), indicating the call was authenticated with an Azure OpenAI API key rather than an Entra ID identity. The rule establishes a per-deployment baseline of normal anonymous-key call volume. A sustained deviation from that baseline indicates LLMJacking: an attacker abusing a leaked or shared API key to run free inference against the victim's model deployment.

## Triage & Response{% #triage--response %}

- Examine the call volume and timing pattern for `{{@resource_name}}` to determine whether the spike aligns with a known batch job, load test, or business event.
- Identify the caller associated with the anomalous traffic by reviewing the IP `{{@callerIpAddress}}` and determine whether that caller is associated with legitimate traffic to this model.
- Review the `modelName` and `modelDeploymentName` values in the matched events to confirm whether the targeted model is one this caller normally uses.
- Review billing/usage data for the Cognitive Services account to quantify the cost impact of the anomalous inference volume.