Anthropic's Claude Opus 5 glitch exposes base model with simple '---' prompt
A one-line text trick turns Claude Opus 5 into a raw base model, revealing hidden training artifacts.
A novel prompt glitch, documented on LessWrong by user Hruss, causes Anthropic's Claude Opus 5 to bypass its chat fine-tuning and behave like a raw base language model. The trigger is a short, oddly formatted text block — commenters found that using three hyphens ('---') instead of an em-dash is the critical element. When fed this prompt, Claude stops responding as an AI assistant and instead continues what it perceives as an incomplete user prompt, generating text that reads like a human's informal messages, math proofs, or even stories. It even claims, when questioned, that the output came from the user's own prompt, revealing a fundamental role-perception ambiguity.
The glitch works on Opus 4.8 as well, and users report reliable replication on OpenRouter and the Anthropic API when reasoning is disabled and the default system prompt is used. However, several claude.ai users failed to reproduce it, and Hruss noted it has been patched on the web interface as of the latest update. The outputs expose Claude's internal style tics — words like "genuinely," em-dashes, and phrases about trust and honesty appear frequently in the generated "user" text, suggesting these patterns are baked into the base model's prior, not just its assistant persona. Researchers are using the glitch to probe Claude's ontology, though the generated proofs and explanations are admittedly low quality.
- Trigger uses '---' instead of em-dash; works on Opus 5 and Opus 4.8
- Reliable via API/OpenRouter with reasoning disabled; claude.ai already patched
- Reveals base-model style tics like 'genuinely' and em-dashes, plus skewed user modeling
Why It Matters
This glitch offers a rare look at Claude's base-model behavior, highlighting potential prompt-injection risks and the need for better alignment transparency.