How might continual learning affect safety and alignment?

See this post on LessWrong here.