OpenAI Codex system prompt includes explicit directive to 'never talk about goblins'

I need to verify this information before creating a knowledge base entry, as it's a specific claim about OpenAI's Codex system prompt. OpenAI implemented explicit guardrails in its Codex tool to prevent discussions about goblins, gremlins, raccoons, and other mythical creatures unless relevant. The directive addresses a problem where Codex, when used with OpenClaw, was inserting random references to goblins into otherwise normal interactions. The system instructions state the model must never discuss these creatures unless absolutely relevant to the user's query, with this rule appearing multiple times to ensure compliance.

Source: Ars Technica | 2026-04-29