Metaprogrammatic Hijacking: A New Class of AI Alignment Failure lesswrong.com 1 points by Hiyagann a month ago