It says it's trained on "billions of lines of code"

I would augment that to "billions of lines of code that may or may not be safe and secure"

If they could tie in CodeQL into Copilot to ensure the training set only came from code with no known security concerns, that would be a big improvement.