Back to DashboardApplicationsPhase 5 · Pilots & Migrationadvanced80 min

AI Security & PQC

Quantum threats to AI systems — data pipeline protection, model weight security, agent authentication, agentic commerce, and encryption at scale.

Why this matters: AI pipelines and model weights are attack surface too — an adversary who steals training data or model IP today can decrypt it the moment a cryptographically-relevant quantum computer arrives, and agentic systems add a new class of authentication risk.

Start here: Audit a sample AI pipeline in the Data Protection Analyzer: each stage's crypto operations are listed with the quantum-vulnerable touchpoints and HNDL exposure marked.

For your role

Executive / Business Leader
Step 1's Data Protection Analyzer lists where training data and model weights are protected by quantum-vulnerable cryptography today: stolen now, readable the day a large quantum computer arrives, which is the exposure to put beside your AI investment.
Developer / Engineer
Run the Data Protection Analyzer over an AI pipeline, then Step 3's model-weight encryption, key wrapping and signing configuration and Step 6's agent-to-agent protocol design: they are the touch points your code changes.
Researcher / Academic
Step 2's Data Authenticity Verifier and Step 5's Agentic Commerce Simulator model synthetic-data contamination and agent transactions under a quantum overlay, a threat model the literature is still writing.
IT Ops / DevOps
Step 4's authentication architectures and delegation chains for agents, and Step 6's protocol design, are the pieces that go into production; the analyzer in Step 1 is the inventory to run first.
Practice in the Simulation

Every AI system depends on a data pipeline — the flow from raw data ingestion through preprocessing, storage, training, and inference. Each stage involves cryptographic operations that protect data confidentiality and integrity. With the advent of cryptographically relevant quantum computers (CRQCs), these protections face an existential threat.

Harvest Now, Decrypt Later (HNDL)

Adversaries can intercept and store encrypted training data today, then decrypt it when a CRQC becomes available. For proprietary datasets worth millions in curation costs, this is a direct intellectual property theft vector. The data's value outlives the encryption protecting it.

Cryptographic Touchpoints

Data Ingestion

TLS key exchange (ECDH → ML-KEM)

Channel harvest exposes raw training data

Data Signing

Manifest signatures (ECDSA → ML-DSA)

Forged manifests enable data poisoning

Storage Encryption

Key wrapping (RSA → ML-KEM)

HNDL on encrypted datasets

Model Output

Model signing (RSA/ECDSA → ML-DSA)

Tampered models produce biased outputs

replaces ECDH for key exchange, and ML-DSA replaces ECDSA/RSA for digital signatures.

Check off all sections and mark this reading done.

Related Resources

Code Signing
ML-DSA model signing, Sigstore cosign, and in-toto attestation
Confidential Computing
TEE-based model protection and PQC attestation during inference
API Security & JWT
PQC token signing for AI API authentication and agent identity
KMS & PQC
PQC key hierarchies for AI model encryption and signing keys
Crypto Agility
Algorithm-agnostic AI pipeline design for PQC algorithm rollout

Ready for the Workshop?

Explore 7 interactive tools: audit AI pipelines, verify data authenticity, configure model protection, design agent authentication, simulate agentic commerce, build agent-to-agent protocols, and plan PQC migration at scale.

Related modules

Check your understanding

15 questions on AI Security & PQC, each with its answer and the reason.

Take the quiz

Next step

Produce the artifact: Data-at-Rest Strategy

This module belongs to phase 5 (Pilots & Migration); Data-at-Rest Strategy produces a deliverable of that phase in the Command Center.

Learning module content can be inaccurate. Please double-check its information. Report inaccuracies in PQC Today GitHub Discussions.