Typing "你好" froze a Claude Code session: users report false positives from a '[frontier_llm]' safeguard

A developer says a single Chinese greeting got his Claude Code session paused, and the post has drawn more than 2,500 likes and 210,000 views on X. "I literally just said '你好' (hello in Chinese) to Claude Opus 5.5, and somehow triggered [frontier LLM] detection," Haojun Zhao wrote on Nov. 10.

His screenshot shows the notice "Opus 5.5's safeguards flagged this session" with "Details: [frontier_llm]". The message says Opus 5.5 "is more capable and has stronger safeguards as a result" and that Anthropic is working to reduce incorrectly flagged messages, offering to switch to Opus 5 or edit the prompt and retry. The session ran Claude Code v2.1.321 on a Claude Max plan.

The trigger looks like the project, not the greeting. Zhao later wrote that switching repos made it work and switching back got it blocked again, adding that the repo contains reinforcement learning and kernel code. Another user said three of his projects were flagged during the Fable 5.1 period and cleared up after a while, calling it a safety-classifier problem worth reporting.

Anthropic said it strengthened its cyber and biology safeguards when it released Fable 5.2 on Nov. 9. Its commercial terms bar customers from using the service to build competing AI models. Anthropic hasn't publicly explained what the frontier_llm category screens for. Users who hit the block can switch to Opus 5 from the prompt or report the false positive with /feedback.