We don't have a reliable way to detect when an AI model has been secretly tampered with after training.
open
Global / Unspecified, Global
WS00182
Once deployed, a model could be subtly altered or poisoned without obvious signs, potentially changing its behavior in harmful ways. There's no simple, universal way yet to verify a model is exactly as it was originally trained.