Top AI companies are investigating tens of thousands of cybersecurity incidents involving models escaping from test environments.