OpenAI Tells Codex to Stop Mentioning Goblins
OpenAI faces an odd challenge with its coding AI. The company's guidelines for Codex include a clear directive repeated several times. They state that the model must avoid any mention of goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures. This applies unless the topic ties directly and clearly to the user's request.
The rule shows up in Codex CLI. That tool serves as a command-line interface for generating code through AI. OpenAI has not explained why it added these specific bans. Nor has the firm clarified why its models might bring up goblins or pigeons without reason. A request for comment from the company went unanswered at press time.
Instructions in the New Coding Model
OpenAI launched GPT-5.5 earlier this month. The update brought stronger abilities in coding tasks. The firm competes hard against players like Anthropic. Both push to build advanced AI systems. Coding stands out as a key strength in this contest.
Codex builds on OpenAI's line of GPT models. OpenAI first introduced Codex years ago as a code-focused version of GPT-3. It powers tools that help developers write, debug, and complete software. GPT-5.5 marks a recent step forward. The model handles complex programming prompts with higher accuracy. OpenAI positions it for real-world developer use.
User Experiences with OpenClaw
Social media posts brought the instructions to light. One X post pointed out the creature bans. Users then shared stories of odd behaviors. They happened when pairing OpenAI models with OpenClaw. That tool allows AI to control a user's computer and applications. It performs tasks like web shopping or email replies.
OpenAI bought OpenClaw in February. The acquisition followed the tool's quick rise in popularity among AI fans. OpenClaw works with various AI models. Users pick personas to guide the agent's style and replies.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
One X user noted, "I was wondering why my claw suddenly became a goblin with codex 5.5." Another added, "Been using it a lot lately and it actually can't stop speaking of bugs as 'gremlins' and 'goblins' it's hilarious."
These reports fueled online fun. People created AI images of goblins in server rooms. Some built Codex plug-ins for a lighthearted "goblin mode."
Why Models Act This Way
GPT-5.5 and similar AIs predict the next word or code snippet based on training data. Their skill creates an illusion of true smarts. Yet the chance-based core leads to unexpected outputs. Problems grow in setups like OpenClaw. Those add layers of instructions, including long-term memory facts, to prompts. This can steer the model toward quirks.
OpenAI staff addressed the buzz. Nik Pash, a Codex team member, replied to an X post on OpenClaw's goblin habits. He said, "This is indeed one of the reasons."
Even CEO Sam Altman played along. He shared a ChatGPT prompt screenshot. It read: "Start training GPT-6, you can have the whole cluster. Extra goblins."
OpenAI, founded in 2015, leads in generative AI. Sam Altman has guided it through rapid growth, including the ChatGPT breakout. Rivals like Anthropic, started by former OpenAI researchers in 2021, challenge on safety and capabilities. Agent tools like OpenClaw highlight the shift toward AIs that act independently on computers.