OpenAI Bans Goblins in GPT-5.5 System Prompt

OpenAI has implemented a specific restriction within the system prompt of its GPT-5.5 model to prevent the AI from discussing certain creatures and animals. The directive, discovered in the open source code for the Codex CLI, instructs the model to 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.

The operational warning became public in April 2026 after OpenAI posted the latest code for the Codex CLI to GitHub. This specific prohibition appears twice within a set of base instructions totaling more than 3,500 words designed to guide the behavior of the recently released GPT-5.5 model.

System Prompt Constraints

System prompts serve as the underlying set of rules that define a model’s persona, safety boundaries, and operational constraints before a user ever submits a query. The instructions for GPT-5.5 include a variety of stylistic and safety guidelines alongside the ban on fantasy creatures.

Among the other directives found in the Codex CLI code are reminders for the model to not use emojis or em dashes unless explicitly instructed. The prompt also includes critical safety guards regarding developer tools, specifically ordering the model to never use destructive commands like ‘git reset –hard’ or ‘git checkout –‘ unless the user has clearly asked for that operation.

The presence of these rules highlights the ongoing effort to refine how AI interacts with technical environments, ensuring that the model does not accidentally execute commands that could result in data loss for a developer.

Addressing Model Behavior

The decision to specifically target creatures like goblins and trolls suggests that OpenAI is addressing a behavioral issue unique to the GPT-5.5 release. Analysis of the JSON file containing the system prompts reveals that instructions for earlier models do not include this specific prohibition.

This discrepancy indicates that the restriction was likely a reactive measure. Anecdotal evidence from social media users has indicated that GPT-5.5 exhibited a tendency to focus on goblins in conversations where the topic was entirely unrelated.

When a large language model develops a penchant for a specific, unrelated topic, it is often a sign of an alignment issue or a recurring pattern in the training data that manifests as an obsession or a hallucination. By hard-coding a prohibition into the system prompt, developers can suppress these behaviors without needing to undergo a full retraining of the model.

The discovery of these instructions provides a rare glimpse into the manual “guardrails” OpenAI uses to stabilize the output of its most recent models, balancing the flexibility of the AI with the need to prevent distracting or irrelevant tangents during technical tasks.