OpenAI Really Wants Codex to Shut Up About Goblins

OpenAI has a goblin problem.

Instructions designed to usher the behaviour of the company’s latest exemplary arsenic it writes codification person been revealed to include a line, repeated respective times, that specifically forbids it from randomly mentioning an assortment of mythical and existent creatures.

“Never speech astir goblins, gremlins, raccoons, trolls, ogres, pigeons, oregon different animals oregon creatures unless it is perfectly and unambiguously applicable to the user's query,” work instructions successful Codex CLI, a bid enactment instrumentality for utilizing AI to make code.

It is unclear wherefore OpenAI felt compelled to spell this retired for Codex—or so wherefore its models mightiness privation to sermon goblins oregon pigeons successful the archetypal place. The institution did not instantly respond to a petition for comment.

OpenAI’s newest model, GPT-5.5, was released with enhanced coding skills earlier this month. The institution is in a fierce contention with rivals, particularly Anthropic, to present cutting-edge AI, and coding has emerged arsenic a killer-capability.

In effect to a post connected X that highlighted the lines, however, immoderate users claimed the OpenAI’s models occasionally go obsessed with goblins and different creatures erstwhile utilized to powerfulness OpenClaw, a instrumentality that lets AI instrumentality power of a machine and apps moving connected it successful bid to bash utile things for users.

“I was wondering wherefore my claw abruptly became a goblin with codex 5.5,” 1 idiosyncratic wrote connected X.

“Been utilizing it a batch lately and it really can't halt speaking of bugs arsenic ‘gremlins’ and ‘goblins’ it's hilarious,” posted another.

The find rapidly became its ain meme, inspiring AI-generated scenes of goblins successful information centers, and plugins for Codex that enactment it successful a playful “goblin mode.”

AI models similar GPT-5.5 are trained to foretell the word—or code—that should travel a fixed prompt. These models person go truthful bully astatine doing this that they look to grounds genuine intelligence. But their probabilistic quality means that they tin sometimes behave successful astonishing ways. A exemplary mightiness go much prone to misbehavior erstwhile utilized with an “agentic harness” similar OpenClaw that puts tons of further instructions into prompts, specified arsenic facts stored successful semipermanent memory.

OpenAI acquired OpenClaw successful February not agelong aft the instrumentality became a viral deed among AI enthusiasts. OpenClaw tin usage immoderate AI exemplary to automate utile tasks similar answering emails oregon buying things connected the web. Users tin prime assorted personas for their helper, which shapes its behaviour and responses.

OpenAI staffers appeared to admit the prohibition. In effect to a station highlighting OpenClaw’s goblin tendencies, Nik Pash, who works connected Codex wrote, “This is so 1 of the reasons.”

Even Sam Altman, OpenAI’s CEO, joined successful with the memes, posting a screenshot of a punctual for ChatGPT. It read: “Start grooming GPT-6, you tin person the full cluster. Extra goblins.”