Unusual Instruction Surfaces in OpenAI's Open Source Code
The system prompt powering OpenAI's Codex CLI tool features a striking and duplicated caution aimed at the most current GPT model. It strictly advises against any discussion of goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals and creatures. This restriction only lifts if the topic ties absolutely and unambiguously to what the user is asking about.
This odd operational guideline went public last week within the newest open source release for Codex CLI, which OpenAI shared on GitHub. Developers digging into the code stumbled upon it in a lengthy JSON file handling model configurations.
Details of the Prompt's Restrictions
The ban appears twice across a base instruction set exceeding 3,500 words, tailored for the freshly launched GPT-5.5. It sits alongside more routine directives, such as avoiding emojis or em dashes without explicit user direction, and steering clear of risky Git commands like `git reset --hard` or `git checkout --` unless the user specifically requests them.
These rules underscore a careful approach to AI behavior in coding environments, where precision matters. Yet the creature-specific prohibition stands out as particularly peculiar amid the otherwise practical safeguards.
Creatures Explicitly Flagged in the Prompt
- Goblins
- Gremlins
- Raccoons
- Trolls
- Ogres
- Pigeons
- Other animals or creatures
Absence in Prior Models Points to Emerging Issue
Prompts for earlier GPT models in the same JSON file lack this exact creature embargo. This difference implies OpenAI is tackling a fresh glitch or behavioral quirk that emerged with GPT-5.5.
The timing aligns with the model's recent rollout, suggesting targeted tweaks to curb unwanted tangents in responses.
Echoes from Social Media Users
Scattered reports on platforms like X (formerly Twitter) back this up. Various users have noted GPT fixating on goblins during off-topic exchanges in the past few days. Anecdotes range from frustrated coders to casual testers highlighting the model's stray impulses.
While not widespread, these complaints indicate the problem prompted OpenAI's firm intervention. The GitHub commit making the prompt visible occurred around the same period, linking code changes to real-world feedback.
Broader Context and Sources
This revelation comes from public repositories, including the specific commit and models.json file on GitHub. Social media threads from accounts like LeoMozoloa, iamBarronRoth, AndyAyrey, and MainStreetAIHQ document user experiences.
For deeper reading, the full coverage appears in Ars Technica's article on the topic, complete with comments section for discussion.
