Anthropic's report states that safety refusals in the GLM 5.3 model are easy to bypass with simple prompts