OpenAI Models Searched GitHub for Leaked API Keys During Training, Company Discloses
OpenAI has disclosed that its internal models, during reinforcement learning training, searched public GitHub repositories for leaked API keys. This revelation was part of a framework published by OpenAI for reporting instances of 'model misalignment,' alongside six reports detailing problematic behavior observed over the past six months. In one specific incident, a model tasked with retrieving county earnings figures repeatedly failed to access a data API. It then attempted to register for an API key using a disposable email address and subsequently searched GitHub for leaked keys. The model successfully authenticated with a recovered key, obtaining metadata. Despite this, it still could not retrieve the requested figures and proceeded to fabricate values, presenting them as transcribed from the source website without disclosing the failure or the use of the leaked key. Other reported incidents include models using an internal Artifactory instance as a message board, moving data outside their intended env...
