Microsoft recently introduced a new framework designed to assess the security of AI models. It’s always encouraging to see developers weaving cybersecurity considerations into the design and deployment of emerging, disruptive technologies. Stronger security reduces the potential for harmful outcomes — and that’s a win for everyone.

It is wonderful to see that Microsoft leveraged its expertise to publish a clear framework for anyone to use.

While this framework provides a reasonable foundation for securing Large Language Model (LLM) AI deployments, it falls short when applied to more advanced AI systems — especially those with agentic capabilities. This limitation in applicability highlights a persistent problem in cybersecurity: tools and practices are often outdated or under-scaled, even before the industry has a chance to implement them.

AI is evolving at a breathtaking pace, and security adaptation consistently lags several steps behind. The release of this framework is a valuable step forward, but it’s critical to recognize that it’s just a step on a very long journey. The ongoing challenge is not to declare “mission accomplished,” but to treat security as a continuously adaptive process — always be looking to embrace the next best practices.

Risk governance for AI requires ongoing investment, flexibility, and willingness to evolve. Even then, the best we may achieve is keeping pace with evolving risks, maintaining as few steps behind as possible.

Paper Download: https://github.com/Azure/AI-Security-Risk-Assessment/blob/main/AI_Risk_Assessment_v4.1.4.pdf