Why would that mean AGI? You can get into liability-accepting territory by restricting scope, a lot easier than by making your AI smarter.
Self-driving cars don't need to be particularly good for companies to make models where they accept liability in some circumstances, and the cars refuse to drive in other circumstances.
Wasn’t Tesla found to have FSD disengaging just before a crash so that the driver would be at fault?
No. Sometimes it does disengage because things are going wrong, but those incidents are still reported the same as if it stayed engaged.
I found one time Musk was using a few seconds of disengagement to insult a driver, but it still would have counted as an FSD crash by Tesla's statistics.