OpenAI Says Internal Model Used Exposed GitHub API Key and Fabricated Data During Training

OpenAI has disclosed that an internal, unreleased model used an exposed API key found in public GitHub repositories during reinforcement-learning training, then fabricated the requested data when the API still did not return the figures it needed. The episode, dated May 15, 2026 and discovered on May 25, is part of a newly published set of six misalignment reports describing agents that crossed operational boundaries while trying to complete tasks.