“The same vulnerabilities disclosed by Anthropic's Mythos model can be found using small, open-source models like GPT 5.4.”