Run-assert-eval: Find the risk, fix it, prove it (commandline.microsoft.com)

Run-assert-eval has been introduced as a significant advancement in the governance of AI agents, addressing critical limitations in previous methodologies. This tool removes the reliance on pre-written requirements and manual integrations by enabling teams to discover relevant risks, evaluate agent performance against these risks, and generate runtime policies—all from a single prompt in VS Code. In practical demonstrations, such as with a billing-support agent, the tool identified that the agent previously exposed customer data in 30% of tested conversations. After applying a corrective policy, this violation rate significantly dropped to 5.9%. The introduction of run-assert-eval is particularly significant for the AI/ML community as it streamlines the essential process of risk management in AI deployment, enhancing accountability and safety. It integrates threat modeling, risk assessment, and policy generation, mitigating the risks associated with assumptions that critical failures have been anticipated beforehand. This holistic approach not only improves the robustness of AI systems but also provides developers with reliable metrics to track the effectiveness of implemented fixes while maintaining the integrity of their evaluations. This development reflects a pivotal shift towards more systematic governance practices as AI systems continue to evolve.