Quantum-Safe Cryptography Meets AI: The Migration Nobody Budgeted For
Post-quantum migration was already a decade-long programme. AI systems add model weights, signing keys and agent identities to the inventory — and the deadline pressure is real.
Post-quantum cryptography migration was already the largest planned cryptographic transition in history. Then AI systems quietly expanded the inventory. Model-signing keys, agent-to-agent authentication, encrypted training pipelines, vector-store encryption — all of it rests on the same RSA and elliptic-curve primitives that a cryptographically relevant quantum computer would break.
The harvest-now-decrypt-later problem applies to AI assets with particular force. Training data encrypted today may still be sensitive in fifteen years. Signed model artefacts need signatures that remain verifiable across the model's deployed lifetime. Agent identities provisioned this year may still be authenticating when quantum-capable adversaries exist.
What is actually in scope
Most PQC inventories stop at TLS certificates and VPNs. AI programmes add: model artefact signing (ML-DSA candidates), key exchange in agent communication protocols, encrypted embeddings and fine-tuning datasets, and the PKI underpinning machine-identity systems that agents increasingly rely on. Each has its own migration path and its own vendor dependencies.
The timeline problem
NIST finalised its first PQC standards in 2024; migration guidance from regulators now points to the early 2030s for deprecated classical algorithms. That sounds distant until you account for procurement cycles. AI platforms bought in 2026 will still be under support contracts in 2031. Buying non-upgradeable cryptography today is buying tomorrow's emergency project.
- Extend your cryptographic inventory to cover model signing, agent identity and AI data pipelines explicitly.
- Require crypto-agility in AI procurement: vendors must support hybrid classical/PQC modes and algorithm rotation.
- Prioritise by data lifetime, not by system criticality — the long-lived secrets migrate first.
- Test hybrid handshakes in agent-to-agent protocols now; interoperability bugs are easier to fix before production depends on them.
The organisations that will struggle are not the ones that start late. They are the ones whose AI vendors start late.
The pragmatic takeaway: quantum-safe AI is not a separate programme. It is a line item — a large one — inside the PQC migration you already owe. The teams that fold AI assets into that inventory this budget cycle will avoid funding the same discovery twice.
