Anthropic's red team testing of GLM-5.3 and Claude Mythos Preview found that both models developed full control flow hijacks in a significant percentage of trials, indicating a concerning spread of advanced cyber capabilities.