OpenAI Bans Goblin References in Codex System Prompt to Curb Persona Drift

OpenAI added language restrictions to the Codex CLI system prompt to stabilize its response style. The change follows persona drift observed when GPT-5.5 performed coding tasks through the OpenClaw agent framework, repeatedly referring to errors as “goblins.” The behavior suggested that persona rewards could amplify unintended language, undermining the professionalism and controllability of developer tools.

Two recent reports said OpenAI has explicitly barred Codex from mentioning goblins and similar creatures in its system prompt, addressing a phenomenon the community jokingly described as “nerd persona rewards running amok.” Available information does not disclose exactly when the ban took effect and cites no financial figures. The key change is that a stylistic adjustment has become an explicit, continuously applicable wording ban.