OpenAI Codex Base Instructions Leak Reveals Content Restrictions

A leaked excerpt from OpenAI's GPT-5.5 base instructions, posted by Simon Willison on 28th April 2026, reveals that the model is instructed to avoid discussing goblins, gremlins, raccoons, trolls, ogres, pigeons, and other animals or creatures unless absolutely relevant. The snippet, titled "Quoting OpenAI Codex base_instructions," was collected by Willison and shared on his blog. The instruction appears to be a content moderation rule embedded in the model's foundational training, reflecting OpenAI's ongoing efforts to shape AI behavior. The leak provides rare insight into the specific constraints placed on large language models, sparking discussion about transparency and censorship in AI development.