From FIPS to FPGA with LLM Agents: Measuring What Tests and Reviews Miss in Post-Quantum Cryptographic Hardware
November 20, 2026
A seven-stage verification-gated workflow ported an ML-KEM-768/ML-DSA-65 accelerator to a Zynq UltraScale+ card using LLM agents for RTL generation. Despite passing NIST ACVP cases, the agents' self-generated simulation gates only caught 76% of 399 tool-selected netlist mutants.
HOW THIS AFFECTS YOU
●
builderDo not rely on LLM-generated tests alone for hardware verification, as they can miss nearly a quarter of faults.
●
researcherThis highlights the gap between passing functional tests and achieving robust verification through mutation testing.