OpenAI 的 Codex 编码指令在 codex-rs/models.json 中被公开，包含一条反复出现的规定：除非与使用者询问内容绝对且明确相关，否则不要谈论 goblins、gremlins、raccoons、trolls、ogres、pigeons，或其他动物与生物。这段内容出现在用于 AI 辅助写程式的 Codex CLI。威尔·奈特（Will Knight）的文章（2026年4月28日晚上7:45）指出，这项政策在 OpenAI 于本月早些时候推出具更强编码能力的 GPT-5.5 后才被揭露，且 OpenAI 未立即回应采访请求。
这条规定之所以引人注意，是因为一些用户报告称，当在 OpenClaw 中使用 GPT-5.5 时，模型会偏向使用大量生物相关语汇。OpenClaw 是一个代理式工具，可让 AI 控制电脑与应用程式完成任务。有人说自己的助理突然变成了 goblin；另一人表示它一直把 bugs 称为 “gremlins” 与 “goblins”。关于这种行为的 X 贴文快速扩散为迷因，催生了 AI 生成的资料中心 goblin 画面，以及能启用玩梗“goblin mode”的 Codex 插件。这种现象虽以搞笑呈现，但也被视为大型语言模型偏态行为的副作用。
OpenAI 先前表示，该问题可能与代理式架构中的提示超载有关：语言模型本质上是机率预测器，来自长期记忆系统的额外指令会放大异常 token 模式。OpenAI 于2月收购 OpenClaw，这是在其在 AI 用户间暴红后不久，并允许使用者为代理行为选择不同人格。对一则投诉贴文回应时，OpenAI 工程师 Nik Pash 称这项生物限制“确实是其中之一”，而 OpenAI 执行长 Sam Altman 也以贴出一则戏仿提示词的方式加入玩梗： Start training GPT-6, you can have the whole cluster. Extra goblins。
OpenAI’s Codex coding instructions, disclosed in codex-rs/models.json, contain a repeated directive: never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant. This appears in Codex CLI, the command-line interface for AI-assisted coding. Will Knight’s article, dated Apr 28, 2026 at 7:45 PM, notes that the policy became visible only after OpenAI released GPT-5.5 with stronger coding ability earlier this month, and that OpenAI did not respond to comment requests right away.
The directive is notable because users reported that GPT-5.5 can drift into creature-heavy language when used in OpenClaw, an agentic tool that lets AI control a computer and apps to complete tasks. Some users said their assistant suddenly became a goblin; another said it kept calling bugs gremlins and goblins. The X post about this behavior quickly became a meme, spawning AI-generated goblin data-center imagery and Codex plugins for a playful goblin mode. The phenomenon, though funny in tone, was also framed as a side effect of large language model behavior.
OpenAI previously argued the issue may relate to prompt overload in agentic setups: language models are probabilistic predictors, and extra instructions from long-term memory systems can amplify unusual token patterns. OpenAI acquired OpenClaw in February, shortly after it went viral among AI users, and allows users to select different personae for agent behavior. In reply to a complaint, OpenAI engineer Nik Pash said the creature ban was indeed one of the reasons, and OpenAI CEO Sam Altman joined the joke by posting a fake prompt line: Start training GPT-6, you can have the whole cluster. Extra goblins.