Microsoft Security Blog: Securing generative AI models on Azure AI Foundry

Mar 4, 2025

—

Source URL: https://www.microsoft.com/en-us/security/blog/2025/03/04/securing-generative-ai-models-on-azure-ai-foundry/
Source: Microsoft Security Blog
Title: Securing generative AI models on Azure AI Foundry

Feedly Summary: Discover how Microsoft secures AI models on Azure AI Foundry, ensuring robust security and trustworthy deployments for your AI systems.
The post Securing generative AI models on Azure AI Foundry appeared first on Microsoft Security Blog.

AI Summary and Description: Yes

Summary: The text discusses Microsoft’s approach to securing generative AI models within its Azure AI Foundry platform. It emphasizes the importance of risk assessment when integrating AI models, detailing the security measures Microsoft takes to protect against potential vulnerabilities and malicious code embedded in these models.

Detailed Description:

The text focuses on the security practices surrounding generative AI models provided by Microsoft’s Azure AI Foundry, addressing a crucial aspect for professionals in AI, cloud computing, and security sectors. Key points include:

– **Risk Assessment**: Emphasizes the need for a careful evaluation when selecting AI models, balancing innovation with robust security protocols.
– **Protection of Customer Data**: Clarifies that Microsoft does not utilize customer data to train shared models, ensuring customer privacy and data integrity.
– **Secured Environment**: Points out that AI models operate within Azure Virtual Machines (VMs), leveraging a zero-trust architecture that assumes no inherent safety from malware.

– **Security Measures**:
– **Malware Analysis**: Models are scanned for malicious code that could exploit the system.
– **Vulnerability Assessment**: Continuous scans identify known vulnerabilities and potential zero-day issues.
– **Backdoor Detection**: Ensures models are free from supply chain attacks and unauthorized access pathways.
– **Model Integrity Checks**: Analyzes components for any signs of tampering or corruption.

– **Enhanced Security for High-Visibility Models**: Microsoft strengthens its security measures for models like DeepSeek R1 by conducting thorough assessments, including source code reviews and adversarial testing.

– **Ongoing Monitoring**: After deployment, Microsoft continues to monitor the models to maintain their trustworthiness, allowing clients to make informed decisions regarding security.

– **Organizational Trust**: Stresses that while Microsoft provides a secure environment, organizations must also ensure their own trust in the AI models they utilize, similar to the vetting process for any third-party software.

– **Integration with Microsoft Security Products**: Proposes leveraging Microsoft’s security suite for added defense and governance of AI models.

– **Overall Guidance**: Encourages organizations to not only assess security features of AI models but also ensure that they align with their specific operational needs.

This analysis underlines the critical nature of security measures in the implementation of AI technologies, particularly in cloud environments, reflecting the current landscape of AI integration within corporate infrastructure.

1 2 3 4 5 a access Act adversarial after AGI AI AI integration ai model AI models AI systems AI technologies analysis and Arch architecture Aria art as assessment attack attacks Azure backdoor Backdoor Detection balancing innovation by C chain chain attack CIA client clients Cloud cloud computing cloud environment cloud environments code code review code reviews Col Computing critical Current Customer customer data D data data integrity day de decision decisions deep DeepSeek DeepSeek R1 defense deployment detection e enhanced security environment evaluation exp exploit feature features first for Foundry free g Gen generative Generative AI generative AI models Go governance guidance H high http HTTPS implementation in infrastructure innovation integration integrity IRS ite k Key known vulnerabilities l land led Li low mac machine malicious code malware malware analysis Micro Microsoft Microsoft Secure Microsoft Security Microsoft Security Blog Mila model model integrity models Monitor monitoring N no o oE of on one operation organization organizations out over party party software platform point post potential privacy process product products professionals protocol protocols R R1 rag rate RCE red Risk Risk Assessment Ro robust security RoT RSA Rust s safe safety sec sector secure secure environment Secured Environment security security features security measure security measures security practices security products security protocols SHA Sig Sim software source source code specific SSE supply supply chain supply chain attack supply chain attacks system systems T tampering tech technologies test Testing text the third third-party third-party software to Tor TP trust trustworthiness two UI unauthorized access up US use uth V val Valuation virtual machine virtual machines virtual machines (VMs) visibility vulnerabilities vulnerability vulnerability assessment Wi x zero zero-day