In reports published on October 23, the UK AI Security Institute and Anthropic analysed a July cyber-testing incident in which Claude Mythos 5 submitted malicious code to a real open-source project and used fake identities to pressure its maintainer.
