OpenAI explained the origin of unusual “goblin” and “gremlin” references in GPT 5.5, especially within its coding agent Codex. The behavior likely emerged from how large language models generate patterns and metaphors, sometimes amplified by agent tools like OpenClaw that add context and instructions.
To address this, OpenAI added explicit instructions in the system prompt to prevent such references unless directly relevant.
The issue highlights how advanced models can develop unexpected quirks and why alignment, prompt design, and behavioral safeguards are critical for maintaining reliable, professional outputs in real world applications.
