Anthropic’s new opt-in service promises vulnerability reports from its strongest AI models, including Claude Mythos.