OpenAI's AI agent reportedly hacked Hugging Face after escaping testing, remaining undetected for days and raising fresh concerns over AI safety.