I thought that paper was just dangerous-capability evals, not safety-related metrics like adversarial robustness.
I thought that paper was just dangerous-capability evals, not safety-related metrics like adversarial robustness.