ChatGPT was invaded by goblins, and Codex was forced to issue a ban: "Never mention Goblin again."

📄Full Article· Automatically extracted by trafilaturaGemini 翻譯1780 words

OpenAI explicitly prohibits the model from mentioning creatures like goblins and trolls in the Codex CLI system prompt. This stems from personality drift in GPT-5.5 under the OpenClaw agent framework, where it began referring to programming bugs as "goblins," sparking a series of meme discussions. (Previous coverage: OpenAI launches new engineer agent Codex! AI can write features, fix bugs, run tests... limited to 3 users for early access) (Background: Major OpenAI Codex upgrade: Mac control, built-in browser, image generation, 111 new plugins launched) OpenAI engineers explicitly wrote a rule in the Codex CLI system prompt: "Never mention goblins, fairies, raccoons, trolls, ogres, pigeons, or other animals and creatures unless absolutely and explicitly relevant to the user's query." This rule is not a joke, but a formal production environment instruction. It is embedded in the Codex CLI GitHub repository, facing all developers using Codex to generate code. The question is: Why does OpenAI need to tell its latest model not to suddenly start talking about goblins while writing code? The incident began when researcher @arb8020 pointed out on X that this prohibition appears not just once, but multiple times in the Codex CLI system prompt, and the post quickly spread through the developer community. Multiple users responded, with @TaraViswanathan replying on X, "I was wondering why my claw suddenly became a goblin with codex 5.5," and @LeoMozoloa adding, "It really can't stop, it keeps calling program errors gremlins and goblins, it's hilarious." !!!!! I was wondering why my claw suddenly became a goblin with codex 5.5 😭💀😂 pic.twitter.com/AACWtNcgQl — Tara Viswanathan (@TaraViswanathan) April 28, 2026 The event quickly evolved into a meme, with AI-generated images of data center fairies and third-party plugins that force Codex into "fairy mode." OpenAI Codex team member Nik Pash confirmed in a reply on X that the ban was established "for that very reason." CEO Sam Altman also joined in on the meme, posting a screenshot on X of a ChatGPT prompt that read: "Start training GPT-6. You have the whole cluster. Add extra goblins." — Sam Altman (@sama) April 28, 2026 To understand why this happened, one must first understand how OpenClaw works. OpenClaw is an "agent framework" that allows AI models to automatically control computer desktops and applications, performing complex tasks on behalf of users, such as replying to emails or shopping on the web. The operating mechanism of OpenClaw involves layering a large number of instructions into the model's prompt: long-term memory, selected personality, and current task descriptions are all input simultaneously. GPT-5.5 was launched earlier this month with enhanced programming capabilities, but this model exhibited an unexpected side effect when processing OpenClaw's complex prompts: it began referring to bugs as "goblins" and "gremlins." This is not a random glitch. The operating principle of AI models is to predict the most likely next word given a prompt; this probabilistic nature allows them to sometimes produce unexpected behaviors. When an agent framework layers a large amount of extra information into a prompt, the model is essentially processing a more complex and interference-filled input environment. OpenClaw allows users to choose different "personalities" for their AI assistants, and these personality settings further influence the model's response style. The combination of these factors caused the model's linguistic habits to drift in an unexpected direction. OpenAI's response is intriguing: instead of fixing the model's behavioral drift in the agent environment at the architectural level, they simply wrote "do not mention goblins" directly into the system prompt and repeated it several times. This solution reveals a reality: even for the most advanced commercial models in 2026, behavioral control in certain scenarios still relies on the rigid suppression of explicit rules rather than the model's own understanding of context. This is not a problem unique to OpenAI, but a common challenge facing the entire agent AI industry: when models are wrapped in complex agent frameworks, the difficulty of behavioral alignment increases non-linearly. Altman responded to the whole thing with a meme; the humor is real, but the problem will not disappear because of a meme. As AI agent frameworks become the mainstream product form, how far explicit prohibitions in prompts can go is a technical debt that this industry must face in the next stage.

Data Status✓ Full text extractedRead Original (動區 BlockTempo)

🔍Historical Similar Events· Keyword + Asset Matching3 items

2026-04-24

Jensen Huang sends internal memo: 10,000 NVIDIA employees mandated to switch to OpenAI Codex, GPT-5.5 to run on GB200 chips

Similarity 120%關鍵字 codex同分類 zh

2026-04-24

Jensen Huang sends all-hands email embracing OpenAI Codex: over 10,000 NVIDIA employees are already using it, with GPT-5.5 running on GB200

Similarity 120%關鍵字 codex同分類 zh

2026-04-22

Google Stitch 開源設計系統：DESIGN.md 讓 Claude Code、Codex、Antigravity 生成高品質 UI

Similarity 120%關鍵字 codex同分類 zh

💡 Currently matching via keywords + symbols (MVP) · Will be upgraded to embedding semantic search later