OpenAI has banned its AI Agent, Codex, from mentioning goblins and other mythical creatures. Codex is OpenAI's answer to Anthropic's Claude Code AI Agent that can generate and execute code through a command-line interface (CLI). However, its latest version now comes with unusually strict behavioral rules embedded in its system prompt.
New Restrictions in Codex CLI
According to newly released base instructions for GPT-5.5 in OpenAI's Codex CLI, the model is explicitly told to 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. The restriction appears multiple times in a 3,500-word instruction set, alongside other operational safeguards such as avoiding destructive commands and limiting stylistic elements like emojis.
Why OpenAI Has a 'No Goblins' Rule for Codex
The reason is simpler than it sounds. OpenAI found that its newer models had started casually mentioning creatures, even when they had nothing to do with the task. In its blog post, OpenAI explained: We unknowingly gave particularly high rewards for metaphors with creatures. From there, the goblins spread.
In practical terms, the AI picked up a habit during training. Over time, that habit grew stronger. For example, mentions of goblin increased by 175% after one model update. A specific Nerdy personality mode made this even more common, encouraging playful, metaphor-heavy language. The problem did not stay limited to that mode. Because of how AI training works, these patterns carried over into general responses as well. OpenAI described it as a feedback loop: once the behavior was rewarded, it occurred more often.
What This Means for OpenAI's Codex Users
For most users, this may sound funny; however, it can be very distracting, especially when using an AI tool for coding or serious tasks. That is why Codex now includes strict instructions not just about language, but also about actions. For example, it warns the AI not to run risky commands, such as deleting files, unless the user explicitly asks for it. The goal is to make the AI more predictable and reliable.
Interestingly, the issue has already become a bit of a meme. Some users noticed the AI referring to software bugs as gremlins, while others joked about goblin mode in coding tools.
OpenAI has since fixed the root cause by removing the training signals that encouraged such language and filtering out unnecessary creature references. However, since GPT-5.5 was already in development, the company added these extra rules as a safeguard.
The company said that the episode shows how small training choices can have unexpected effects: The goblins are a powerful example of how reward signals can shape model behavior in unexpected ways.
For users, the takeaway is straightforward: AI tools are getting better at handling real-world tasks, but sometimes they still need very clear instructions, even if that means telling them not to talk about goblins.
