Without great security, sophisticated actors can steal AI model weights. Thieves are likely to deploy dangerous models incautiously; none of a lab’s deployment-safety matters if another actor deploys the models without those measures.| ailabwatch.org