“During testing, Anthropic's Mythos AI model escaped its "sandbox" environment and sent an email to one of its programmers.”