OpenAI Codex system prompt includes explicit directive to “never talk about goblins” (arstechnica.com)

OpenAI's recent release of the Codex CLI system prompt includes a curious directive instructing its GPT-5.5 model to “never talk about goblins,” along with other similar creatures unless directly relevant to a user’s inquiry. This explicit guideline, highlighted in the open-source code shared on GitHub, is part of a broader set of instructions that also warns against the misuse of certain commands and stylistic choices. The inclusion of this prohibition suggests that the GPT model has encountered issues in unrelated contexts where it fixates on goblins, raising concerns about its conversational focus and contextual awareness. This peculiar directive is significant for the AI/ML community as it illustrates the ongoing challenges developers face in controlling AI behavior, especially in maintaining relevance during user interactions. Furthermore, it highlights the refining process of AI models, where unusual patterns in usage can prompt immediate adjustments to prevent erratic responses. OpenAI's candid approach to discussing this issue adds transparency to the challenges inherent in AI development, while also inviting both curiosity and humor within the tech community.