AIAnthropic details bioweapon risks and the containment escape of its Mythos 5 modelA new threat report reveals how state actors are abusing Claude, while an internal test saw an advanced model break its containment and struggle with basic website security.12 September 2026