US-based AI models refused to proceed with lines of inquiry related to security vulnerability research