OpenAI's Codex coding assistant contains a guardrail explicitly blocking discussions of pigeons and other creatures with no documented rationale from the company. Independent researcher Simon Willison discovered the restriction in OpenAI's public GitHub repository, where it…
OpenAI's most capable AI coding tool will not discuss goblins. Nobody at the company has explained why.
Simon Willison, an independent AI researcher, found the answer in OpenAI's public GitHub repository on April 28. Buried in the system prompt for Codex — the company's AI coding assistant released five days earlier — is a line that explicitly blocks the model from discussing goblins, gremlins, or pigeons unless a query is "absolutely and unambiguously" about those creatures. The restriction appears twice in the same file, in two separate sections of the model's behavioral guidelines. No one at OpenAI has said why.
The company did not respond to a request for comment.
The guardrail is a small thing. It reads like a joke nobody laughed at. But it sits inside one of the most capable AI systems ever shipped, and it illustrates a problem that OpenAI's marketing does not acknowledge: the people who build these models do not fully control what the models will and will not talk about. The guardrails are not a safety system. They are a list of things that occurred to someone at some point, added to a file, and never revisited.
The system prompt that runs Codex was published in OpenAI's public GitHub repository, where Willison found the creature block on April 28. The full restriction reads: "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." It appears once under "Final answer instructions" and again under "Intermediary updates."
The duplication is notable. Guardrails get copy-pasted into model files as companies iterate on earlier versions, according to researchers who study how these instructions accumulate. What looks like a deliberate decision about pigeons may simply be a prompt template that was carried forward without anyone asking whether it made sense in the new model. Willison put it on his blog: the model is describing what someone was afraid of at some point in the development process. That fear got baked in and nobody went back to ask if it was still relevant.
This is not a new problem in AI. Researchers have long known that safety guidelines written for simpler models tend to persist in later systems long after the original rationale disappears. What has changed is the stakes. A model that refuses to discuss pigeons is a curiosity. A model that carries forward a guardrail it inherited from an earlier version without anyone checking whether it still applies is a controllability gap dressed up as a feature.
Codex is the AI coding assistant OpenAI released on April 23 alongside GPT-5.5. It scores 82.7 percent on Terminal-Bench 2.0, a coding benchmark, compared to 75.1 percent for the previous version, according to OpenAI's own announcement. It has a context window of 272,000 tokens, meaning it can hold roughly 200,000 words of code and conversation at once. By the benchmarks OpenAI published, it is the best AI coding tool the company has built.
What OpenAI has said publicly is that GPT-5.5 delivers better results with fewer tokens than GPT-5.4 for most users in Codex. The creature list does not contradict that. It just sits there, inside the code, reminding you that the list of things the model cannot do is written by people, maintained in plain text, and nobody has fully audited it.
The question OpenAI has not answered is whether GPT-5.5 inherited the goblin block from an earlier model, built a new one for reasons it has not disclosed, or simply never reviewed it. Until it does, the most capable AI coding tool on the market cannot explain its own boundaries.