Output Length Effect on DeepSeek-R1's Safety in Forced Thinking