OpenAI bans Codex from talking about goblins and pigeons in its AI agent instructions

A strange rule has surfaced in the system instructions for Codex, OpenAI's new AI programming tool: the model is explicitly forbidden from mentioning goblins, gremlins, raccoons, trolls, ogres, pigeons and other creatures unless the user's request is related to them. Judging by the reaction of developers and users, the rule appeared after a very real bug in agent mode.

What was found in the instructions

The trigger was a line in the instructions for Codex CLI, the command-line tool that uses the model to generate and edit code. The wording leaves little room for interpretation: the agent is told not to talk about "goblins, gremlins, raccoons, trolls, ogres, pigeons, and other animals or creatures" unless this is "absolutely and unambiguously" related to the request. Users noticed that the restriction is repeated several times, so this is not a stray note but a deliberately codified behavioral rule.

At first glance it looks like an inside joke by the team, but the context is quite serious. OpenAI has just doubled down on programming: GPT-5.5 shipped with improved coding capabilities, and competition with Anthropic and other players for the market in AI tools for developers has intensified sharply. Against that backdrop, any oddity in model behavior stops being a harmless curiosity and becomes a product problem, especially when the agent works not only in a chat but also controls real applications.

Where the goblins came from

The ban seems to have a history. After screenshots of the instructions spread across X, users began recalling that OpenAI models paired with OpenClaw did sometimes fixate on this vocabulary. One developer wrote that his claw "suddenly became a goblin" after switching to Codex 5.5. Another noted that the agent kept calling bugs "goblins" and "gremlins." In an ordinary chatbot this would look like a quirky manner of speech, but in a coding agent it is noise that makes the output harder to understand and trust.

"Now I see why my claw suddenly became a goblin with

Codex 5.5."

WIRED links this to how modern models behave inside agent harnesses. A base model predicts the next token and usually stays within the bounds of the task, but in agent mode the prompt is supplemented with memory, system instructions, a set of roles and service hints. The longer and more complex this stack, the higher the chance that the model latches onto recurring metaphors, random associations or an inappropriate style. If the AI is also controlling a computer, answering emails or placing orders, even a minor verbal mutation signals that the loop needs tight constraints.

How a meme became a product matter

The story quickly spilled out of engineering chats and became a meme. Users began posting generated scenes of goblins in data centers, and Codex even got plugins with a playful "goblin mode." More important, though, OpenAI employees effectively confirmed that the ban is tied to real model behavior. Codex developer Nik Pash wrote in response to the discussion that it was "indeed one of the reasons." In other words, the company is not just joking; it is closing off a specific class of behavioral failures.

• The screenshot of the instructions spread quickly across social media
• Users began sharing similar cases in OpenClaw
• Plugins and memes around "goblin mode" appeared for Codex
• Sam Altman joined the wave with a joke about "extra goblins" in training GPT-6

OpenClaw's role is a separate nuance. The tool lets users 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 went viral, so model behavior inside that interface is no longer external exotica for the company but part of its own platform. The deeper OpenAI moves toward autonomous agents, the more important it becomes to suppress not only dangerous errors but also odd shifts in speech.

What this means

The goblin episode looks like an anecdote, but it shows something more important: developers of AI agents increasingly have to fix not only factual errors but also the model's style of reasoning in complex scenarios. For the coding-assistant market, it is a signal that the competition has moved to the level of fine-tuning behavior. An agent's reliability today depends not only on the quality of the code it writes, but also on how predictably it speaks and acts under load.