Glitch: Specific Prompts Cause Claude Opus to Leak Chain of Thought

matthen2 · x · 2026-07-30

A user discovered that sending specific prompt formats (like “see the below —”) to Claude 3 Opus triggers an unusual response.

Instead of replying normally, the model leaks its internal antml thinking syntax and hallucinates a user request to complete. This unexpected behavior offers an interesting case for studying LLM safety boundaries and internal mechanisms.

Related event: Specific Prompts Trigger Claude Opus Anomalies(2 posts)→

Original post →

More from Fun

Fun channel →