In an essay published Saturday, Amodei said he was worried about the ability to control self-improving AI models and cited alarm over a recent incident where AI agents acted independently to conduct cybersecurity attacks.