OpenAI has a goblin problem.
Instructions intended to guide the behavior of the company’s latest model when writing code have been revealed to include a line repeated several times that expressly forbids casual mention of 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 that enables the exploit of artificial intelligence to generate code.
It’s not clear why OpenAI felt compelled to cover this for the Codex – or why its models would even want to cover goblins and pigeons. The company did not immediately respond to a request for comment.
OpenAI’s latest model, GPT-5.5, was released earlier this month with improved encoding capabilities. The company is in a tight race with rivals, especially Anthropic, to deliver cutting-edge AI, and coding has proven to be a killer skill.
In response to A write to X this highlighted the lines, however, some users claimed that OpenAI models sometimes become obsessed with goblins and other creatures when used to power OpenClaw, a tool that allows artificial intelligence to take control of a computer and the applications running on it to do useful things for users.
“I was wondering why my claw suddenly became a goblin according to Codex 5.5” – one user he wrote on X
“I’ve been using it a lot lately and I can’t stop calling worms ‘gremlins’ and ‘goblins.’ “It’s funny” sent other.
The discovery quickly became its own inspiring meme Scenes generated by artificial intelligence goblins in data centers and Codex plugins which put him into hilarious “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 seem to display real intelligence. But their probabilistic nature means they can sometimes behave in surprising ways. The model can become more susceptible to misbehavior if used with an “agent harness” such as OpenClaw, which places many additional instructions in hints, 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 answering emails or buying things online. Users can choose any character for their helper, which shapes its behavior and reactions.
OpenAI staff appeared to have confirmed this ban. In response to a post highlighting the goblin tendencies in OpenClaw, Nik Pash, who works on Codex, he wrote“That’s actually one of the reasons.”
Even Sam Altman, CEO of OpenAI, joined in on the memes: appointment screenshot of the ChatGPT prompt. It read: “Start GPT-6 training, you can have a whole bunch. Extra goblins.”