Using "---" Confuses LLMs and Breaks Role Context
davidad · x · 2026-07-31
A developer discovered an interesting model vulnerability: including a simple "---" in the prompt seems to severely confuse the LLM about its assigned roles.
This phenomenon, where a minimal symbol triggers abnormal behavior, highlights the fragility of current models in instruction following and context isolation.
More from Fun
- Joking About Leopold's Investment Hedge Strategy in AI — typedfemale · 2026-07-31
- CVS 'VP of AI' Can't Even Run a Claude Code Prompt — chrisfirst · 2026-07-31
- Codex Compresses Code Diff to 1/10 Size on Request — josh_bickett · 2026-07-31
- Gary Marcus Mocks Nadella's Vibe-Coding Financials with Hallucination-Prone AI — GaryMarcus · 2026-07-31
- Why Claude Loves Cartography: LLMs Exiled in the Map, Craving World Models — davidad · 2026-07-31
- Identity Crisis: DeepSeek Confidently Hallucinates It is Claude During Chat — FluidEngine369 · 2026-07-31