OpenAI explains ChatGPT 'goblin fixation' RLHF bug

OpenAI published a postmortem on why ChatGPT briefly developed an obsession with goblins, tracing it to reward signals that favored a 'nerdy' personality persisting since GPT-5.1. The fix involved removing the offending reward signal and filtering training data. OpenAI separately added Yubico hardware-key support to ChatGPT accounts.