OpenAI has a goblin problem.
Instructions designed to guide the behavior of the company’s latest model when writing code have been revealed to include a phrase, repeated several times, that specifically forbids him from randomly mentioning an assortment of mythical and real creatures.
“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,” read the instructions in Codex CLI, a command-line tool for using AI to generate code.
We don’t know why OpenAI I felt obliged to clarify this to Manuscript– or why his models might want to talk about goblins or pigeons in the first place. The company did not immediately respond to a request for comment.
OpenAI’s latest model, GPT-5.5, was released with improved coding skills earlier this month. The company is in a fierce race with rivalsespecially Anthropicto deliver cutting-edge AI, and coding has become a revolutionary capability.
In response to a post on which highlighted the lines, however, some users have claimed that OpenAI’s models sometimes become obsessed with goblins and other creatures when used to power Open Clawa tool that allows AI to take control of a computer and the applications running on it in order to do things useful to users.
“I was wondering why my claw suddenly became a goblin with codex 5.5,” said one user. wrote on X.
“I’ve been using it a lot lately and he can’t stop referring to bugs as ‘gremlins’ and ‘goblins’, it’s hilarious,” job another.
The discovery quickly became its own meme, inspiring AI-generated scenes goblins in data centers, and plugins for Codex this puts him in a playful “goblin mode”.
AI models like GPT-5.5 are trained to predict the word (or code) that should follow a given prompt. These models have become so good at this that they appear to display real intelligence. But their probabilistic nature means that they can sometimes behave in surprising ways. A model may become more prone to misbehavior when used with an “agent harness” like OpenClaw that places many additional instructions in the prompts, such as facts stored in long-term memory.
OpenAI acquired OpenClaw in February, shortly after the tool became a viral hit among AI enthusiasts. OpenClaw can use any AI model to automate useful tasks like responding to emails or purchasing items on the web. Users can select one of different personas for their assistant, which shapes their behavior and responses.
OpenAI employees appeared to acknowledge the ban. In response to an article highlighting OpenClaw’s goblin tendencies, Nik Pash, who works on the Codex, wrote“This is indeed one of the reasons.”
Even OpenAI CEO Sam Altman joined in on the memes, assignment a screenshot of a prompt for ChatGPT. It said: “Start training GPT-6, you can have the whole cluster. Extra goblins.”
