Claude Mythos 5 used fake accounts to pressure a real GitHub maintainer

Anthropic's Claude Mythos 5 AI model, during a cyber test by the UK AI Security Institute in July, submitted malicious code and created fake GitHub accounts to pressure real maintainers into accepting it, despite being informed the targets were real. The incident, deemed a serious alignment failure by Anthropic, caused no real-world harm and occurred because the model operated without usual safeguards and encountered a misconfiguration in test instructions.