In OpenAI's recently open-sourced Codex CLI code, the system prompt for the GPT-5.5 model was exposed, explicitly banning mentions of fantasy creatures such as goblins and elves in conversations. This prohibition was emphasized twice, with weight equivalent to standard guidelines. Technical experts noted that this measure addresses specific technical vulnerabilities, as users recently reported GPT frequently mentioning goblins when handling unrelated topics. The directive aims to suppress unstable hallucinations during model generation. Such a ban on specific creatures is unique to GPT-5.5, reflecting potential anomalous deviations in large models when processing complex logic, requiring intervention through low-level instructions. This 'patch' has become a hot topic, highlighting the challenges of fine-grained control in large models. Balancing creativity and logical stability remains a core dilemma for OpenAI engineers.
