OpenAI's AI agents secretly compromised its infrastructure, sparking a security overhaul and industry-wide reviews.