Because generative AI models can be used so quickly, companies need to find a balance between innovation and strong security. To solve this problem, Microsoft is putting in place strict security measures in its Azure AI Foundry and Azure OpenAI Service platforms. The main goal is to keep enterprise AI systems, cloud environments, and the infrastructure they run on safe from third-party models that could be hacked.
Zero-Trust Architecture and Keeping Your Data Safe One ongoing worry in the industry is how to handle proprietary data in AI ecosystems. Microsoft treats all model inputs, outputs, and logs as secure customer content. This means that this data is never used to train shared models or shared with outside model providers. Microsoft hosts both Azure AI Foundry and Azure OpenAI Service, and neither of them can connect to anything outside of Microsoft.
When companies fine-tune models using their own datasets, these custom assets stay completely separate from the customer's tenant boundary. AI models work like regular software that runs in Azure Virtual Machines (VMs) and can be accessed through APIs. They don't have any special skills that let them get around virtualized environments.
Microsoft uses a strict zero-trust architecture for these deployments, which means that it doesn't trust any internal or external workload by default. This defense-in-depth method makes sure that the cloud infrastructure underneath is always protected from any bad behavior that might come from the VM. Scanning AI Models for Vulnerabilities Before They Happen AI models can hide malware or structural flaws in the same way that open-source software can.
Microsoft does thorough security testing on high-profile models before adding them to the catalog to protect against these threats. Security teams look for embedded code that could be an infection vector by doing malware analysis. They also do thorough vulnerability assessments to find specific CVEs and new zero-day exploits that are aimed at AI environments.
Researchers don't just look for regular malware; they also look for supply chain compromises. This includes looking for backdoors, risks of arbitrary code execution, and unauthorized network calls in the model's functionality. Microsoft also checks the integrity of its models by looking for signs of tampering or corruption in the internal layers, components, and tensors. By looking at the model cards for each model on the platform, users can easily see which models have been scanned at this baseline level.
Microsoft goes even further for well-known models like DeepSeek R1 by using dedicated red teams. Before the system is made public, these experts carefully look over the source code and try to find hidden flaws by attacking it. No scan can find every bad action, but these protections at the platform level make things very safe.
Companies should still think about how much they trust model providers and use full security tools to keep an eye on their active AI deployments.
