Open-weight AI models are catching up to the frontier. The safety gap remains.

Open-weight AI models, which release their underlying code for anyone to download and run, are now approaching the performance level of frontier models created by leading AI companies. However, a significant safety gap has emerged: while frontier models include protections like refusal training and classifiers to prevent misuse on dangerous tasks like cyberattacks or biological engineering, open-weight models can be run without these safeguards once downloaded. This matters because once someone obtains the model weights, protections become unenforceable and can be removed or modified. The challenge for policymakers and developers is finding ways to limit dangerous capabilities in open-weight models while preserving their beneficial uses, since techniques like training data filtering work better for some domains than others.