The episode argues cryptography is undergoing an identity migration: a field built on hiding information has pivoted to confronting systems that can't be made to hide—where the threat lives inside the machinery rather than on the wire. The central, recurring insight is that "the optimization is the vulnerability"—efficiency tricks like federated adapters, quantization, and speculative decoding become attack surfaces precisely because they make systems cheaper, faster, or more practical. The proposed response is to extend cryptography's oldest rigorous instincts—formal verification, certificate-bound authority, policy enforcement—over probabilistic, non-deterministic AI agents, attempting to impose mathematical certainty on systems designed never to be certain.

Topics: LLM security, federated learning backdoors, quantization attacks, formal verification, autonomous agents