Generative AI is transforming many industries, but businesses often struggle with how to create and deploy safe and secure AI tools as technology evolves. Leaders worry about the risk of AI generating incorrect or harmful information, leaking sensitive data, being hijacked by attackers or violating privacy laws — and they’re sometimes ill-equipped to handle the risks.  

“Organizations care about safety and security along with quality and performance of their AI applications,” says Sarah Bird, chief product officer of Responsible AI at Microsoft. “But many of them don’t understand what they need to do to make their AI trustworthy, or they don’t have the tools to do it.”  

To bridge the gap, Microsoft provides tools and services that help developers build and ship trustworthy AI systems, or AI built with security, safety and privacy in mind. The tools have helped many organizations launch technologies in complex and heavily regulated environments, from an AI assistant that summarizes patient medical records to an AI chatbot that gives customers tax guidance.  

The approach is also helping developers work more efficiently, says Mehrnoosh Sameki, a Responsible AI principal product manager at Microsoft. 

This post is part of Microsoft’s Building AI Responsibly series, which explores top concerns with deploying AI and how the company is addressing them with its responsible AI practices and tools.

“It’s very easy to get to the first version of a generative AI application, but people slow down drastically before it goes live because they’re scared it might expose them to risk, or they don’t know if they’re complying with regulations and requirements,” she says. “These tools expedite deployment and give peace of mind as you go through testing and safeguarding your application.”  

The tools are part of a holistic method that Microsoft provides for building AI responsibly, honed by expertise in identifying, measuring, managing and monitoring risk in its own products — and making sure each step is done. When generative AI first emerged, the company assembled experts in security, safety, fairness and other areas to identify foundational risks and share documentation, something it still does today as technology changes. It then developed a thorough approach for mitigating risk and tools for putting it into practice.  

The approach reflects the work of an AI Red Team that identifies emerging risks like hallucinations and prompt attacks, researchers who study deepfakes, measurement experts who developed a system for evaluating AI, and engineers who build and refine safety guardrails. Tools include the open source framework PyRIT for red teams to identify risks, automated evaluations in Azure AI Foundry for continuously measuring and monitoring risks, and Azure AI Content Safety for detecting and blocking harmful inputs and outputs.  

Microsoft also publishes best practices for choosing the right model for an application, writing system messages and designing user experiences as part of building a robust AI safety system.  

“We use a defense-in-depth approach with many layers protecting against different types of risks, and we’re giving people all the pieces to do this work themselves,” Bird says. 

For the tax-preparation company that built a guidance chatbot, the capability to correct AI hallucinations was particularly important for providing accurate information, says Sameki. The company also made its chatbot more secure, safe and private with filters that block prompt attacks, harmful content and personally identifiable information.