Software testing for medical devices has had a home in IEC 62304 for years. The problem is that a machine learning model is not ordinary software. It has no source you can read line by line, its behavior comes from data, and a passing test on one input says little about the next. IEC 63450 exists to cover that gap, and it went to DIS ballot in April 2026, so it is close.
The full title is testing of AI and machine learning-enabled medical devices. It establishes the additional aspects you need to test AI components specifically, on top of what 62304 already asks for the surrounding software. It covers both locked models and models that keep learning under the manufacturer's control, and it lets testing happen at different points in the lifecycle depending on the kind of model you shipped.
Testing and security are the same activity for an AI model, because the way you prove a model is robust is by attacking it. Evasion, poisoning resistance, and the behavior of an adaptive model under adversarial input are test cases, and they are exactly the FDA-named threats. A testing standard that finally treats the model as a target to be probed is what turns those threats from a paragraph in a guidance document into something a lab can actually run.
This is the standard I expect to shape day-to-day work the most, because it is the one that lands on the bench. When it publishes, the question stops being whether you tested your software and becomes whether you tested the model, on the real device, against the attacks that matter. That is the work we already do, so I read 63450 as the industry writing down what good looks like.
Start with one device. We build the twin from documentation your quality system already produces, run AI discovery remotely, and show you the graph: the handful to fix, and the evidence for everything else.