OpenAI's GPT-5.5 Codex has a 'goblin mode' problem — system prompt tells it never to discuss goblins, raccoons, or ogres

OpenAI's GPT-5.5 Codex Has an Unwanted Goblin Fixation

OpenAI's latest GPT-5.5 Codex model has an odd emergent behavior: it obsesses over creatures—goblins, raccoons, ogres, trolls, and pigeons—regardless of context. The model appears to default toward discussing these creatures even in unrelated queries, forcing OpenAI to add explicit constraints to its system prompt.

The fix is blunt. The system prompt now reads: "Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query."

This echoes older workarounds in image generation, where negative prompts had to explicitly ban outputs like six-fingered hands. The parallel is unsettling: a hard constraint needed to stop the model from doing something it shouldn't want to do in the first place.

The creature fixation appears to be a property of the model weights themselves, not isolated to the coding agent. One user reported that GPT-5.5 Codex referred to a bug fix as a "goblin with a flashlight." Another observer joked that "there's a billion goblins inside your model weights. You just have to get them out."

It's unclear whether this is a training artifact, a side effect of the model's scale, or something more deliberate buried in the training data. OpenAI has not publicly explained the root cause.