LLM agent hardware flow ports ML-KEM and ML-DSA accelerator, but tests miss many injected faults
Researchers describe a seven-stage verification-gated workflow that ported a post-quantum cryptography accelerator to FPGA hardware, then audited how much its own tests and reviews actually catch.
In plain words
The authors used AI agents to help design and check a chip implementing NIST's post-quantum encryption and signature algorithms, running on an FPGA board. They then measured how well the flow's own tests and AI reviews caught deliberately injected faults, finding significant gaps that formal mathematical checks closed.
What was announced
In an arXiv paper abstract, the authors present a seven-stage verification-gated workflow, developed during a post-quantum design port, in which a test gate counts only if it can fail on negative controls and fault mutants, and reviewer agents audit work before commit. The flow ported an ML-KEM-768/ML-DSA-65 accelerator to a 300 MHz Zynq UltraScale+ key-custody card, with all six operations byte-exact on the board, 85 of 285 applicable NIST ACVP cases passing, and 484 ML-DSA known answers passing.
The audit shows passing module tests overstate verification strength. Across three RTL modules, existing simulation gates killed only 76% of 399 tool-selected netlist mutants; tests derived from the specification and guided by surviving mutants raised the fresh-draw decapsulation kill rate from 48% to 74%. In pre-registered single-model replays, undirected whole-design review caught two verifier flaws in only 3/6 and 1/6 sessions, while module-scoped review caught every known flaw in all 54 sessions. Yet even scoped review detected planted key-custody flaws in only 3/20 sessions, whereas formal checks flagged every one. The authors conclude that trustworthy agentic hardware needed author-independent mutants for test gates, module-level scoping for LLM review, and formal checks for custody invariants.
What it does not show
- The evidence is the paper's abstract only; full methodology, experimental setup, and peer review status are not visible in the supplied passages.
- All reported results, including the port, kill rates, and review session outcomes, come from the authors' own audit; no independent confirmation is available in the supplied material.
- The abstract reports results for three RTL modules and specific flaw replays; broader applicability to other designs is not established in the supplied text.
For security and IT teams
- Scope
- The reported port targets an ML-KEM-768/ML-DSA-65 accelerator on a 300 MHz Zynq UltraScale+ key-custody card, with results specific to that board and design.
Sources
Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 21 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.