Building production-ready solutions with agentic AI comes with inherent risks. When agents make mistakes or hallucinate, the potential impacts can multiply rapidly.
“It turns out that it’s very easy to write AI-powered software, but it’s very hard to write AI-powered software that works right in real-world cases,” says Yonatan Zunger, CVP and deputy CISO for Microsoft.
Yunger explains how important it is to test if you want to build trustworthy agentic AI.
Learn from our experience
Read our practical advice about applying security fundamentals to AI.
Key takeaways
Here are best practices to apply while building trustworthy agentic AI:
- Prototype. Test. Iterate. Think of and try prompts your real users might give your agentic AI. Use real data. From those trials, build a set of test cases and keep testing.
- Use AI tools to amplify testing. Evaluating agents requires a “try it and repeat it” mindset. Using AI Foundry with such tools as Python Risk Identification Tool amplifies these assessment capabilities.
- Record your tests. Applying this practice, as you would with unit testing, enables you to repeat evaluations as your data models and agents evolve.
- Don’t skimp on testing. Test early, test often, test with real data. This is the best way to understand what your agent might do when it encounters the unexpected.
Try it out
Related links
- Read more about how to secure agentic AI.
- Learn more from Zunger about how to deploy AI safely.
- Discover what you need to know about governing autonomous agents.
- Explore how to use Azure AI Foundry to secure generative AI models.
- Generate synthetic data for fine-tuning in Microsoft Foundry.
- Leverage built-in policies for model deployment in Microsoft Foundry.
