MIT's blog post: https://news.mit.edu/2026/new-method-enables-ai-safety-criti...