OpenAI Implements Guardrails to Stop Codex Goblin References

OpenAI implemented new system guardrails for its Codex coding assistant to prevent the AI from randomly discussing goblins, gremlins, and other animals. The company discovered that GPT-5.5, which powers Codex, frequently used terms like goblin as casual substitutes for things or software bugs. This behavior was particularly prevalent on OpenClaw, an agentic AI platform OpenAI acquired.

Investigations revealed the issue originated from a Nerdy personality setting used during training, which incentivized creature-based metaphors to achieve a playful tone. Although the Nerdy setting accounted for only 2.5% of queries and was retired in March, reinforcement learning caused these style tics to spread to general responses. Mentions of goblins reportedly increased by 175% following the launch of GPT-5.1 in November.

To resolve the glitch, OpenAI filtered training data, removed problematic reward signals, and added specific instructions to the Codex CLI forbidding mentions of creatures unless they are unambiguously relevant to a user query. CEO Sam Altman and engineer Nick Pash acknowledged the quirk on social media, with Altman jokingly referring to it as a goblin moment. Following the viral reaction, Altman announced an invite-only event for developers in San Francisco on May 5 to celebrate the GPT-5.5 release.