Large language model weights represent billions of dollars of training compute. The cryptographic signing and encryption protecting these artefacts is vulnerable to quantum attack. This workshop maps the exposure and builds a migration plan.
For MLOps engineers and AI platform security teams. Covers quantum cryptographic exposure of model weights, training checkpoints, NCCL/MPI cluster communications, model registry signing, and PQC migration using FIPS 203/204/205 for distributed AI training infrastructure.
Training a frontier LLM costs tens of millions of dollars in GPU compute. The resulting model weights are signed with ECDSA or RSA, stored encrypted in cloud object stores, and transmitted across multi-node GPU clusters using NCCL AllReduce or MPI collectives. Every one of these cryptographic operations is vulnerable to a quantum computer running Shor's algorithm. The harvest-now-decrypt-later threat is particularly acute for model weights: an adversary recording encrypted checkpoint files today could decrypt them once quantum capability arrives, gaining access to proprietary model architectures and training data encoded in the weights. This workshop maps the full cryptographic dependency chain of a distributed training pipeline, from data ingest through GPU cluster communications to checkpoint storage and model registry. We then build a phased PQC migration plan using FIPS 203 (ML-KEM) for key encapsulation, FIPS 204 (ML-DSA) for model signing, and FIPS 205 (SLH-DSA) for long-lived provenance attestation, addressing the specific performance constraints of GPU cluster communications and large-artefact signing.
Full-day session structure with scheduled breaks. Content is configurable to your training infrastructure, GPU cluster architecture, and model deployment pipeline.
| # | Session | Topics |
|---|---|---|
| 1 | AI Training Infrastructure and Its Cryptographic Dependencies Where cryptography protects model weights, checkpoints, and cluster communications | |
| 2 | Model Artefact Security Under Quantum Threat Weights, checkpoints, and model registry integrity |
|
| Break, after 50 min | ||
| 3 | Distributed Training Cluster Communications NCCL, MPI, and GPU-to-GPU cryptographic exposure |
|
| 4 | Interactive Demonstration: Model Security Cryptographic Audit Full-day format only |
|
| Break, after 60 min | ||
| 5 | PQC Migration for AI Training Infrastructure FIPS 203/204/205 applied to ML workflows |
|
| 6 | Compliance and Vendor Landscape NIST AI RMF, EU AI Act, and platform readiness |
|
| 7 | Q&A and Migration Planning | |
Workshops are designed and delivered by QSECDEF in collaboration with sector specialists. All facilitators have direct experience in both quantum technologies and AI training infrastructure security.
Workshop design and delivery
QSECDEF brings world-leading expertise in post-quantum cryptography, quantum computing strategy, and defence-grade security assessment. Our advisory membership spans 600+ organisations and 1,200+ professionals working at the intersection of quantum technologies and critical infrastructure security.
Domain expertise and operational validation
AI workshops are co-delivered with sector specialists who bring direct operational experience in distributed training systems, GPU cluster security, and MLOps pipelines. This ensures workshop content is grounded in the operational realities of production AI training infrastructure.
Sessions are configured around your training infrastructure, GPU cluster architecture, model registry tooling, and existing security controls. Get in touch to discuss requirements and schedule a date.
Contact Us