OpenAI Codex system prompt includes explicit directive to "never talk about goblins"

Key Points:

• OpenAI's Codex CLI system prompt for GPT-5.5 includes a repeated explicit warning to avoid mentioning goblins, gremlins, raccoons, trolls, ogres, pigeons, or similar creatures unless absolutely relevant, reflecting a new issue with the latest model.
• The prohibition is unique to GPT-5.5's instructions and was made public on GitHub, with social media users reporting that the model had been inappropriately focusing on goblins in unrelated conversations.
• OpenAI staff deny the warning is a marketing stunt, though some executives have humorously referenced the "goblin moment," and there is discussion of potentially adding a "goblin mode" toggle to override the restriction.
• This situation mirrors a prior issue with xAI’s Grok model, which briefly exhibited problematic behavior due to unauthorized system prompt modifications, prompting increased transparency around system prompts.
• The Codex system prompt also encourages the model to exhibit a vivid inner life, balancing intelligence and playfulness to create a warm, curious, and collaborative interaction experience that feels like engaging with a distinct subjectivity rather than a mere tool.