OpenAI Really Wants Codex to Shut Up About Goblins
Codex CLI's secret rule: no goblins, gremlins, or raccoons allowed
OpenAI's Codex CLI, a command-line tool for AI-powered code generation, has revealed a peculiar prohibition in its system instructions: the model must never discuss goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless absolutely relevant to the user's query. This rule, repeated multiple times, aims to curb unexpected behavior in GPT-5.5, the latest model released earlier this month with enhanced coding skills. The company is locked in a fierce race with rivals like Anthropic to deliver cutting-edge AI, and coding has emerged as a killer capability.
The issue appears linked to OpenClaw, a tool OpenAI acquired in February that lets AI control computers and apps. When used with OpenClaw's agentic harness, which adds extensive instructions and long-term memory to prompts, GPT-5.5 sometimes fixates on goblins and similar entities. Users reported the model speaking of bugs as 'gremlins' and 'goblins,' spawning memes and plugins for 'goblin mode.' OpenAI staffer Nik Pash acknowledged the link, and CEO Sam Altman joined the fun, posting a mock prompt for GPT-6 training with 'extra goblins.'
- Codex CLI instructions explicitly ban mentions of goblins, gremlins, raccoons, trolls, ogres, and pigeons unless relevant
- GPT-5.5 models using OpenClaw's agentic harness spontaneously discuss these creatures, sparking user reports and memes
- CEO Sam Altman acknowledged the issue with a joke about including 'extra goblins' in GPT-6 training
Why It Matters
Highlights AI models' probabilistic quirks and the challenges of controlling agentic tools in production