1Protecting Frontier Model WeightsFrontier model weights are the copyable core of a model’s capability, so protecting them requires layered security rather than one perfect lock.Weights and Provenance 2The AI Supply Chain for Eval ArtifactsModels, datasets, adapters, and benchmarks are not harmless files; they are dependencies that deserve supply-chain controls.Weights and Provenance 3Malware-Lab Discipline for AI EvalsRisky AI evaluations should be run like malware detonations: disposable, isolated, credential-poor, and watched from end to end.Weights and Provenance 4Hermetic Builds, Provenance and SLSABuild integrity depends on sealed inputs, trustworthy records, and a shared standard that applies to software, models, and datasets alike.Weights and Provenance 5Benchmark Contamination and Answer VaultingA benchmark only measures capability if the model cannot recall or retrieve the answers it is supposed to earn.Weights and Provenance