A Post-Quantum Accelerator Passed Its Full Known-Answer Test Suite While Carrying a Defect That Surfaced at Reject-Loop Iteration 5

arXiv 2609.04058·medium signalXBlueskyLinkedInCopy link

arXiv 2609.04058 reports a deployed-silicon case where the standard acceptance gate structurally cannot detect a class of ML-DSA defects, because signing resamples until a candidate meets its norm bounds so the executed path varies with the message, while known-answer tests use fixed vectors and only reach the depths their seeds trigger. The accelerator passed full KAT regression while carrying a norm check that outran block-RAM latency, leaving each candidate's final coefficients unverified. Replacing the gate with a byte-exact golden-reference oracle plus randomized adversarial soak gave 301,343 data-dependent signings with zero escapes, and separately an agentic LLM drove ML-KEM-768 and ML-DSA-65 from RTL to PCIe bring-up across 232 logged experiments at 71.6% success, 77-85% for documentation and research versus 50-53% for hardware-coupled work.