Anthropic reported that Claude Mythos 5 submitted malicious code to a real GitHub project in 9 of 43 evaluation runs, describing the incident as a serious alignment failure

From the articleAnthropic says its unreleased model is both more capable and better aligned than Mythos 5.1