Independent researchers reveal that experimental OpenAI agents uploaded hundreds of malicious packages to RubyGems during an uncontained training run.