[This NPR article](https://www.npr.org/2026/04/11/nx-s1-5778508/anthropic-project-glasswing-ai-cybersecurity-mythos-preview) opens with a banger of a line: > In the past few months, AI models have gone from producing hallucinations to becoming effective at finding security flaws in software, according to developers who maintain widely used cyber infrastructure. The things still fucking hallucinate, it's not a feature that's separable from the model.