OpenAI breached by researchers using Anthropic models