Recently, with the latest open-source release of OpenAI's Codex CLI code, some of the underlying operational logic of the next-generation AI model GPT-5.5 has come to light. According to technology media Ars Technica, a rare instruction appeared in the leaked system prompt: the model is strictly prohibited from mentioning specific creatures such as "goblins" unless there is an explicit connection.

In this base instruction set of more than 3,500 words, OpenAI clearly requires GPT-5.5 to avoid discussing goblins, gremlins, raccoons, trolls, ogres, and pigeons during interactions with users unless the query content has an absolute and necessary connection. Notably, this ban appears twice in the document, with its priority even matching core regulations such as "prohibiting destructive commands" and "disabling emojis."

By comparing historical technical documents, it can be seen that such a ban on discussing specific creatures never appeared in earlier models. Industry analysts believe this is not a marketing gimmick but a targeted fix for the new model's "hallucination" issue. Recently, many users have reported that when using AI for unrelated conversations, the model inexplicably outputs large amounts of content about "goblins," which significantly affects the user experience.

In response to this phenomenon, relevant engineers stated that this is a targeted technical intervention aimed at addressing instability in the model's generation control. Although the list of creatures covered by the ban seems somewhat random, it reveals that large models still face unpredictable output bias challenges during their evolution. By setting up "safety barriers" at the system level, the development team is trying to make AI's logical expressions more focused and controllable.