OpenAI’s instructions for its Codex coding agent include a line — repeated several times — explicitly forbidding the model from mentioning goblins, gremlins, raccoons, trolls, ogres, pigeons, or other creatures “unless it is absolutely and unambiguously relevant to the user’s query.” The unusual directive was discovered in Codex CLI, a command-line tool that uses AI to generate code, and quickly spread across social media in April 2026.
The reason appears to be rooted in real user complaints. Following a post on X highlighting the prohibition, multiple users reported that OpenAI’s models had developed a habit of describing software bugs as “gremlins” and “goblins” when used through OpenClaw, a computer-automation tool OpenAI acquired in February 2026. “Been using it a lot lately and it actually can’t stop speaking of bugs as ‘gremlins’ and ‘goblins’ it’s hilarious,” one user wrote.
Nik Pash, an OpenAI staffer who works on Codex, appeared to confirm the connection, writing in response to a post about OpenClaw’s goblin tendencies that the creature ban was “indeed one of the reasons.”
The behavior likely stems from how large language models work. Models like GPT-5.5 — OpenAI’s newest release, which launched with enhanced coding capabilities earlier this month — predict the most probable next word or piece of code given a prompt. When paired with an “agentic harness” like OpenClaw, which loads additional instructions and memory into prompts, models can behave in unexpected ways.
OpenClaw, which went viral among AI enthusiasts before OpenAI acquired it, allows any AI model to automate tasks such as answering emails or making web purchases. Users can also assign the tool different personae, which shapes how it responds.
The discovery spawned its own meme cycle, including AI-generated images of goblins in data centers and Codex plug-ins that activate a “goblin mode.” Even OpenAI CEO Sam Altman joined in, posting a mock ChatGPT prompt that read: “Start training GPT-6, you can have the whole cluster. Extra goblins.”
Source: WIRED