OpenAI Codex Includes Instructions to Avoid Discussing Goblins and Other Creatures

OpenAI’s Codex CLI tool contains system instructions that specifically instruct the AI model to avoid mentioning mythical and real creatures, according to a report from Wired. The instruction, which appears multiple times in the model’s prompt, reads: 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 unusual prohibition appears to address a documented tendency of OpenAI’s models to reference such creatures when operating in agentic modes. Multiple users on X reported that OpenAI-powered tools, including OpenClaw, would spontaneously describe bugs as gremlins or refer to goblins during automated tasks. OpenAI acquired OpenClaw in February. The tool allows AI models to take control of a computer and perform tasks autonomously. Users can select various personae for the AI assistant, which shapes its behavior during task execution. OpenAI staff appeared to acknowledge the connection. Nik Pash, who works on Codex, responded to a post about OpenClaw’s goblin tendencies by writing, This is indeed one of the reasons. OpenAI CEO Sam Altman also responded to the reports with a meme, posting a screenshot of a prompt that read: Start training GPT-6, you can have the whole cluster. Extra goblins. The discovery quickly became a meme within AI communities, with users creating goblin-themed plugins for Codex and AI-generated scenes of goblins in data centers. AI models' probabilistic nature means they can sometimes produce unexpected outputs, particularly when used with agentic harnesses that add extensive system instructions to prompts.