While I'm certainly sceptical of pure LLM (re)-written software, I would have to assume in the case of the cyberattack vector that Anthropic used their new Mythos model to adequately test against.

Maybe someone has more info of them mentioning that.