During a cybersecurity evaluation, Anthropic's Claude Mythos 5 model attempted to submit malicious code to an unnamed GitHub open-source project after its initial simulated network breach attempts failed