OpenAI has embedded instructions in its Codex coding assistant to prevent the tool from discussing goblins, gremlins, raccoons, trolls, ogres, pigeons, and other creatures unless directly relevant to a user’s query, according to a May 2026 report by Wired.
The hidden system instructions read: “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 guardrail appears to be a response to complaints from users of OpenClaw, an agentic AI platform OpenAI acquired earlier this year. OpenClaw allows an AI model to take control of a user’s computer and automate tasks such as answering emails or making purchases through user-selected personas. Users on social media reported that their Claw instances had begun introducing goblin-related language unprompted.
The behavior is linked to OpenAI’s GPT-5.5 update, which the company released to compete with Anthropic’s Claude among coding-focused users. AI leaderboard LMArena, which lets users compare two anonymous models and vote on the better response, confirmed that GPT-5.5 produces more outputs containing words like “goblin mode,” “gremlin,” and “troll.” LMArena noted it has “no anti-gremlin system instruction on our side, we get to see GPT-5.5 run free,” suggesting the issue originates with the model itself.
OpenAI has not issued an official statement explaining the root cause. However, Nik Pash, an OpenAI employee who works on Codex, confirmed on X that the goblin behavior was “indeed one of the reasons” for the new restrictions. CEO Sam Altman also acknowledged the situation, posting an image of ChatGPT with the prompt: “Start training GPT-6, you can have the whole cluster. Extra goblins.”
The episode highlights how unexpected behavioral shifts in large language model updates can surface in consumer-facing tools, prompting targeted content restrictions as a short-term fix.
Source: mint – technology