OpenAI Really Needs A Codex To Cover Goblins

OpenAI has a goblin problem.

Guidelines designed to guide the practices of the latest generation of companies in the way they code have been disclosure plus a line, repeated several times, that specifically forbids the mention of different types of fictional and real creatures.

“Don’t talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it’s relevant to the user’s question,” read instructions in the Codex CLI, a tool that uses AI to generate code.

It is not known why OpenAI he was forced to explain this Codex– or why his models would want to discuss squirrels or pigeons first. The company did not immediately respond to a request for comment.

The latest version of OpenAI, GPT-5.5, was released with improved coding capabilities earlier this month. The company is in fierce competition with competitorsespecially Anthropicto provide more expensive AI, and copying has emerged as a deadly force.

In response to a post on X which highlighted the lines, however, some users reported that OpenAI models sometimes get busy with goblins and other creatures when used. OpenClawa tool that allows AI to monitor computers and the programs running on them to perform useful actions for users.

“I was wondering why my claw suddenly turned into a goblin with codex 5.5,” another user he wrote on X.

“I’ve been using it recently and it can’t stop talking about bugs like ‘gremlins’ and ‘goblins’ is funny,” he said. has been sent someone.

What he found was his own, inspiring Features created by AI of goblins in the data center, and Codex plugins which puts it in the “goblin” game.

AI models like GPT-5.5 are trained to predict words – or codes – that should follow quickly. These models have been so successful that they seem to have real intelligence. But their abilities mean that they can sometimes do extraordinary things. The model can be difficult to misbehave when used with a “burner” such as OpenClaw that installs additional instructions, such as values ​​stored in long-term memory.

OpenAI acquired OpenClaw in February shortly after the tool became a hit among AI enthusiasts. OpenClaw can use any type of AI to perform basic tasks such as answering emails or shopping online. Users can choose any of the different people to help them, who create their own systems and solutions.

OpenAI staff seem to agree with the ban. In response to the revelations about OpenClaw’s tendencies, Nik Pash, who works on Codex, he wrote“That’s one of the reasons.”

Even Sam Altman, CEO of OpenAI, joined the memes, sending Detailed image of ChatGPT. He said: “Start training GPT-6, you can have the whole team.