OpenAI Banned Codex from Talking About Goblins and Pigeons in AI Agent Instructions

In the system instructions of Codex, OpenAI's new AI tool for programming, an unusual rule was found: the model is directly forbidden from mentioning goblins, gremlins, raccoons, trolls, ogres, pigeons and other creatures if it is not related to the user's request. Based on the reaction of developers and users, the rule appeared after a quite real bug in agent mode.

What was found in the instructions

The impetus was a string in the Codex CLI instructions — a command-line tool that uses a model to generate and edit code. The formulation leaves no room for imagination: the agent is instructed not to talk about "goblins, gremlins, raccoons, trolls, ogres, pigeons and other animals or creatures" unless it is "absolutely and unambiguously" related to the request. Moreover, as users noted, this restriction is repeated several times, meaning it is not a random insertion, but a consciously established behavioral rule.

At first glance, this looks like an internal team joke, but the context of the story is quite serious. OpenAI has just increased its stakes in programming: GPT-5.5 was released with improved coding capabilities, and competition with Anthropic and other players for the AI-tools-for-developers market has intensified sharply. Against this backdrop, any strangeness in the model's behavior ceases to be a harmless curiosity and turns into a product problem, especially if the agent not only works in chat but also manages real applications.

Where the goblins came from

It appears the ban did not come out of nowhere. After screenshots of the instruction were shared on X, users began recalling that OpenAI models in conjunction with OpenClaw indeed sometimes got stuck on such vocabulary. One developer wrote that his claw "suddenly became a goblin" after switching to Codex 5.5. Another noted that the agent constantly called bugs "goblins" and "gremlins". For a regular chatbot this would look like a strange manner of speech, but for a coding agent — like noise that interferes with understanding the output and trusting it.

"Now it's clear why my claw suddenly became a goblin with

Codex 5.5".

WIRED attributes this to how modern models behave inside agent frameworks. The base model predicts the next token and usually stays within task bounds, but in agent mode, memory, system instructions, a set of roles, and service prompts are added to the prompt. The longer and more complex such a chain becomes, the higher the chance that the model will latch onto repetitive metaphors, random associations, or inappropriate style. If the AI is managing a computer, answering emails, or processing purchases, even a minor verbal mutation becomes a signal that the loop needs strict constraints.

How a meme became a product

The story very quickly went beyond the engineering chat and turned into a meme. Users began posting generated scenes with goblins in data centers, and plugins with a game-like "goblin mode" even appeared for Codex. But more importantly: OpenAI employees actually confirmed that the ban is related to the model's actual behavior. Codex developer Nick Pash, in response to the discussion, wrote that this was "really one of the reasons". That is, the company is not just joking, but closing a specific class of behavioral failures.

• Screenshots of the instruction quickly spread across social media
• Users began sharing similar incidents in OpenClaw
• Plugins and memes with "goblin mode" appeared around Codex
• Sam Altman picked up the wave with a joke about "extra goblins" during GPT-6 training

A separate nuance is the role of OpenClaw. This tool allows you to connect almost any model to a computer, give it access to applications, and choose different personas for the assistant. OpenAI acquired OpenClaw in February shortly after the project's viral rise, so the behavior of models within such an interface is no longer external exotica for the company, but part of its own platform. The deeper OpenAI goes into autonomous agents, the more important it becomes to suppress not only dangerous errors, but also strange speech shifts.

What this means

The goblin story may seem like an anecdote, but in fact it shows something more important: AI-agent developers increasingly treat not only factual errors, but also the model's thinking style in complex scenarios. For the coding-assistant market, this is a signal that the battle is already being fought at the level of fine-tuning behavior. An agent's reliability today is determined not only by the quality of code it writes, but also by how predictably it communicates and acts under load.