Security for Azure AI services
Security should be considered a top priority in the development of all applications, and with the growth of artificial-intelligence-enabled applications, security is even more important. This article outlines various security features available for Azure AI services. Each feature addresses a specific liability, so multiple features can be used in the same workflow.
For a comprehensive list of Azure service security recommendations see the Azure AI services security baseline article.
Security features
| Feature | Description |
|---|---|
| Transport Layer Security (TLS) | All of the Azure AI services endpoints exposed over HTTP enforce the TLS 1.2 protocol. With an enforced security protocol, consumers attempting to call an Azure AI services endpoint should follow these guidelines:
|
| Authentication options | Authentication is the act of verifying a user's identity. Authorization, by contrast, is the specification of access rights and privileges to resources for a given identity. An identity is a collection of information about a principal, and a principal can be either an individual user or a service. By default, you authenticate your own calls to Azure AI services using the subscription keys provided; this is the simplest method but not the most secure. The most secure authentication method is to use managed roles in Microsoft Entra ID. To learn about this and other authentication options, see Authenticate requests to Azure AI services. |
| Key rotation | Each Azure AI services resource has two API keys to enable secret rotation. This is a security precaution that lets you regularly change the keys that can access your service, protecting the privacy of your service if a key gets leaked. To learn about this and other authentication options, see Rotate keys. |
| Environment variables | Environment variables are name-value pairs that are stored within a specific development environment. Environment variables are more secure than using hardcoded values in your code. For instructions on how to use environment variables in your code, see the Environment variables guide. However, if your environment is compromised, the environment variables are compromised as well, so this isn't the most secure approach. The most secure authentication method is to use managed roles in Microsoft Entra ID. To learn about this and other authentication options, see Authenticate requests to Azure AI services. |
| Virtual networks | Virtual networks allow you to specify which endpoints can make API calls to your resource. The Azure service rejects API calls from devices outside of your network. You can set a formula-based definition of the allowed network, or you can define an exhaustive list of endpoints to allow. This is another layer of security that can be used in combination with others. |
| Data loss prevention | The data loss prevention feature lets an administrator decide what types of URIs their Azure resource can take as inputs (for those API calls that take URIs as input). This can be done to prevent the possible exfiltration of sensitive company data: If a company stores sensitive information (such as a customer's private data) in URL parameters, a bad actor inside that company could submit the sensitive URLs to an Azure service, which surfaces that data outside the company. Data loss prevention lets you configure the service to reject certain URI forms on arrival. |
| Bring your own storage (BYOS) | The Speech service doesn't currently support Customer Lockbox. However, you can arrange for your service-specific data to be stored in your own storage resource using bring-your-own-storage (BYOS). BYOS allows you to achieve similar data controls to Customer Lockbox. Keep in mind that Speech service data stays and is processed in the Azure region where the Speech resource was created. This applies to any data at rest and data in transit. For customization features like Custom Speech and Custom Voice, all customer data is transferred, stored, and processed in the same region where the Speech service resource and BYOS resource (if used) reside. To use BYOS with Speech, follow the Speech encryption of data at rest guide. Microsoft doesn't use customer data to improve its Speech models. Additionally, if endpoint logging is disabled and no customizations are used, then no customer data is stored by Speech. |
Next steps
- Explore Azure AI services and choose a service to get started.