Close to 12,000 valid secrets that include API keys and passwords have been found in the Common Crawl dataset used for training multiple artificial intelligence models.