The Secret Compartment Inside an AI Model
A new cryptographic construction shows how an adversary could hide an encrypted neural network inside a host model and activate it with a secret trigger. The result exposes a limit of AI capability audits: failing to elicit a behavior does not prove it is absent.