OpenAI has a goblin problem.<\/p>\n
Instructions designed to guide the behavior of the company\u2019s latest model as it writes code have been revealed to include<\/a> a line, repeated several times, that specifically forbids it from randomly mentioning an assortment of mythical and real creatures.<\/p>\n \u201cNever 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,\u201d read instructions in Codex CLI, a command line tool for using AI to generate code.<\/p>\n It is unclear why OpenAI<\/a> felt compelled to spell this out for Codex<\/a>\u2014or indeed why its models might want to discuss goblins or pigeons in the first place. The company did not immediately respond to a request for comment.<\/p>\n OpenAI\u2019s newest model, GPT-5.5, was released with enhanced coding skills earlier this month. The company is in a fierce race with rivals<\/a>, especially Anthropic<\/a>, to deliver cutting-edge AI, and coding has emerged as a killer-capability.<\/p>\n In response to a post on X<\/a> that highlighted the lines, however, some users claimed the OpenAI\u2019s models occasionally become obsessed with goblins and other creatures when used to power OpenClaw<\/a>, a tool that lets AI take control of a computer and apps running on it in order to do useful things for users.<\/p>\n \u201cI was wondering why my claw suddenly became a goblin with codex 5.5,\u201d one user wrote<\/a> on X.<\/p>\n \u201cBeen using it a lot lately and it actually can’t stop speaking of bugs as \u2018gremlins\u2019 and \u2018goblins\u2019 it’s hilarious,\u201d posted<\/a> another.<\/p>\n The discovery quickly became its own meme, inspiring AI-generated scenes<\/a> of goblins in data centers, and plugins for Codex<\/a> that put it in a playful \u201cgoblin mode.\u201d<\/p>\n AI models like GPT-5.5 are trained to predict the word\u2014or code\u2014that should follow a given prompt. These models have become so good at doing this that they appear to exhibit genuine intelligence. But their probabilistic nature means that they can sometimes behave in surprising ways. A model might become more prone to misbehavior when used with an \u201cagentic harness\u201d like OpenClaw that puts lots of additional instructions into prompts, such as facts stored in long-term memory.<\/p>\n OpenAI acquired OpenClaw in February not long 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 on the web. Users can select various personas for their helper, which shapes its behavior and responses.<\/p>\n OpenAI staffers appeared to acknowledge the prohibition. In response to a post highlighting OpenClaw\u2019s goblin tendencies, Nik Pash, who works on Codex wrote<\/a>, \u201cThis is indeed one of the reasons.\u201d<\/p>\n