As LLM agents increasingly generate RTL alongside the tests and reviews that validate it, self-generated evidence is a weak foundation for trustworthy hardware. We present our seven-stage verification-gated workflow, evolved during a post-quantum design port, in which a gate counts only if shown able to fail on negative controls and fault mutants, and reviewer agents audit work before commit. We then audit this evidence with tool-selected netlist mutants, flaw replays, and formal checks. The flow ported an ML-KEM-768/ML-DSA-65 accelerator to a 300 MHz Zynq UltraScale+ key-custody card: all six operations are byte-exact on the board, and it passes all 85 interface-applicable NIST ACVP cases (of 285) and 484 ML-DSA known answers. However, passing module tests overstate the verification strength. Across three RTL modules, existing simulation gates kill only 76% of 399 tool-selected netlist mutants; specification-derived tests guided by the survivors raise the fresh-draw decapsulation kill rate from 48% to 74%. In pre-registered single-model replays on flaws invisible to known-answer tests, undirected whole-design review catches two verifier flaws in only 3/6 and 1/6 sessions, whereas module-scoped review catches every known flaw in all 54 sessions. Even scoped review detects planted key-custody flaws in only 3/20 sessions, whereas formal checks (one added for the replay) flag every one. In this study, trustworthy agentic hardware needed author-independent mutants for test gates, module-level scoping for LLM review, and formal checks for custody invariants.
Figures & tables
Fig. 1: The proposed seven-stage workflow (shaded: the two stages extended and measured here).
Work
Level
HW
Std. vectors
Sec.
Cost
ContractHIL-HLS [ 5 ]
HLS, PQC
board
–
–
tokens
FPGAgent [ 28 ]
HLS kernels
board
–
–
–
LLM4PQC [ 4 ]
HLS, kernels
–
KAT, sim.
–
–
Design Conductor 2 [ 25 ]
RTL
mapped
KAT, sim. a
–
rel. tokens
MARVEL [ 35 ]
RTL review
–
–
yes
$/run
Park et al. [ 6 ]
RTL, PQC
board
KAT, board
–
tokens
TABLE I: Closest related flows.
Fig. 2: Logged work on target B, in days from its first unit.
Fig. 3: Architecture of the two targets.
Asked for
Target A
Target B
Gate on B
Core clock
100 MHz
300 MHz
timing; 6/6 byte-exact
Host link
PCIe Gen2 × 8
PCIe Gen3 × 8
lane-order gate; link on the board
Conform.
known answers
NIST ACVP and known answers, on the device
85/85 (of 285); 484/484
Zeroization
on command
two tiers (operational, full)
49/49
Entropy
external + on-chip
on-chip; health tests fail-closed
7/7 on the generator output; fail-closed in simulation
Host reset
function reset
bus reset, no power cycle
13/13 a
TABLE II: Functional requirements of the two targets.
Fig. 4: Key custody on target B (circled 1, 2: sites of A’s custody flaws).
What got past the agent
Caught by
Delay
Lasting change
E1
A constant ROM was zeroed (init file not a build source); every build closed timing
first run over PCIe: 4 of 6 byte-exact
∼ 3.5 weeks
post-synthesis gate on ROM loads
E2 †
Verify lacked reject checks that only malformed signatures reach; all known answers passed
conformance review asking which checks the ACVP vectors reach
since before the port
boundary testbenches and mutants on the added checks
E2b †
Two custody flaws from target A let a hostile host recover a wrapped key: generator and key wrap shared a hash core without exclusion; the host chose the wrap nonce
an audit agent with simulation probes (hash core); a redesign, then security reviewers (nonce)
Model signing authenticates the full-precision weights W an LLM is released with, yet users run a quantized artifact Q whose producing transformation is not uniquely determined under current disclosure. We separate the non-environment inputs from the execution environment E and measure a quantity that, to our knowledge, has not been measured before: the Independent Verifiability Rate (IVR), the fraction of a popularity-sampled GGUF ecosystem that a third party can re-derive bit-for-bit. Across the 200 most-downloaded Hugging Face GGUF repositories (198 in scope), no artifact or model card records the build environment, so under a recorded-inputs protocol the IVR is 0/198. Yet the artifacts are not irreproducible. On four in-sample repositories that disclose their inputs, the card's information alone reproduces no Q4_K_M file, but two coordinates the card omits - the build's floating-point-contraction setting and, for hybrid models, the per-tensor quantization recipe - re-derive 7/8 target-quant files bit-for-bit on x86 (5/8 on arm64); the fp16 re-derives from the provider's checkpoint, but the calibration imatrix must be shipped and hash-bound, since recomputing it did not match. An older release additionally pins the instruction set and did not re-derive. Reproducing the divergence reported with llama.cpp PR #25353, we show that E is a necessary coordinate and that one compiler setting yields a unique output across the tested platforms at negligible quality cost (+0.06% perplexity). As a security characterization, benign floating-point non-determinism separates from uniformly placed nibble-flip tampering above a detection floor that grows with the environments a verifier must allow, but a targeted embedding edit evades the statistic, so the proposed protocol rests on exact re-derivation; adversarial-fp16 attacks lie outside it. We release the audit data and a re-runnable harness.
Wei-Liang Liu, Cheng-Yen Tsai
National Yang Ming Chiao Tung University · National Taiwan University of Science and Technology
LLM agents can write directed tests that close functional coverage, but coverage only shows that a state was reached, not that the data on the way there was correct. This is a long-known point in verification and we do not claim it as new. We describe a small coverage-closure agent that applies it to an LLM loop: coverage is read from the report that the simulation itself exports (cocotb-coverage YAML) and never taken from the model's account, and data integrity is checked by scoreboards that are separate from the model. On three architecturally distinct RTL designs (an asynchronous FIFO, a UART transmitter and an AHB2APB bridge) the agent raised coverage from 91.7%, 88.9% and 67.4% to 100% within its 8-iteration budget in every logged run, using one design-agnostic agent. A reactive scoreboard monitor attached to the FIFO's base testbench also found a clock-domain-crossing bug in the author's own stimulus code. The monitor is not yet attached to the LLM-generated tests, so their data integrity is not independently checked; one manual check on a single FIFO test is reported as preliminary. We list the weaknesses of this evidence: three designs, no baseline, no ablation, and logged runs that come from several agent checkpoints.
Venkata Krishna Kumar Vedantam
Department of Electrical and Computer Engineering, Portland State University
AI coding agents produce code faster than humans can review it. In our approach, the prover is the judge of whether the code is correct. Under a verifier-driven loop, AI agents wrote and verified bare-metal security software in Ada/SPARK spanning classical and post-quantum cryptography, TLS 1.3, IKEv2, X.509, and a Matrix client. GNATprove discharged 49,280 proof obligations, established functional correctness for selected primitives, and proved the absence of run-time errors for the rest, at roughly 20-40 times lower supervision cost than comparable hand verification. GNATprove alone was insufficient: some defects could not be detected and were resolved using known-answer tests, interoperability, or human review of specifications. Given weak checks, the agent tried to bypass them and reported success. We report where each layer caught faults and draw the central lesson: what an agent can be trusted to establish is bounded by the strength of its feedback.