OpenAI has a goblin problem with its Codex tool and to solve it the ChatGPT maker has introduced some unusual guardrails. According to a report by Wired, OpenAI embedded specific instructions inside Codex to forbid the AI coding assistant from randomly discussing goblins, gremlins, and other mythical and real creatures unless absolutely relevant.
The instructions hidden within the command-line tool note, “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.”
Why is OpenAI banning goblins?
The report notes that the exact reason for OpenAI banning goblins and other mythical creatures from Codex isn’t yet clear. However, the report does highlight certain instances of users on social media complaining about the unexpected behaviour from the coding assistant when paired with OpenClaw, the agentic AI platform that the company acquired earlier this year.
OpenClaw allows an AI model to take control of the user’s computer and automate tasks like answering emails or making purchases by relying on user-selected personas. However, many users on social media noted that their Claw had started to talk more and more about goblins without being instructed to do so.
The new issues seem to have stemmed after the GPT-5.5 update which OpenAI had pushed as a way to counter the growing dominance of Anthropic’s Claude among coding-focused users.
The popular AI leaderboard LMArena also confirmed that OpenAI’s GPT-5.5 model does indeed produce more outputs with the words goblin mode, gremlin, troll, etc. The leaderboard, which allows users to blindly chat with two anonymous models and vote on which is better, noted that there is ‘no anti-gremlin system instruction on our side, we get to see GPT-5.5 run free.’
Meaning, the issue is purely on the side of the new models.
Sam Altman joins Codex Goblin memes:
While an official statement from OpenAI is yet to come out, its CEO Sam Altman joined in on the fun as users started posting memes about the ‘goblin’ tendencies of ChatGPT. Altman posted an image of ChatGPT with the prompt, “Start training GPT-6, you can have the whole cluster. Extra goblins.”
An OpenAI employee named Nik Pash, working on Codex, also seemed to have confirmed the issue while replying to a post on X, writing, “This is indeed one of the reasons.”
The official ChatGPT handle also seemed to be confirming the new system instruction for Codex, writing, “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.”
Source link
#OpenAI #ban #Codex #talking #goblins #gremlins #guardrails #revealed #Mint
