Trust, but Validate the Instrument: Auditing AI-Generated RTL Verification Plans on Authored Security-Regression Proxies

TL;DR

SecTB-RTL framework audits AI-generated RTL verification plans, revealing mismatches in generation and execution rules.

cs.CR 🔴 Advanced 2026-09-17 7 views
Hang Xiao Chuhong Xu Kainan Zhou Gangzhen Qian Lu Yi
hardware security RTL verification large language models mutation testing trustworthy AI

Key Findings

Methodology

The SecTB-RTL framework covers 31 tasks and 124 hardware security regressions, using a deterministic non-AI baseline for comparison. Task-level studies ensure generated verification plans remain unchanged in the trusted production path. The framework is designed to ensure that each generated verification plan is evaluated under standard simulation, synthesis, and formal tools.

Key Results

  • In the C1-R3 run, only 9 out of 1,860 calls passed production semantic validation, highlighting mismatches in generation and execution rules.
  • The deterministic baseline killed 36, 75, and 78 mutants at increasing resource limits, demonstrating benchmark feasibility.
  • Generated verification plans may fail to detect security regressions even with high structural coverage.

Significance

This study highlights the limitations of AI-generated RTL verification plans in hardware security, emphasizing the importance of consistency between generation and execution. It provides a methodological foundation for future AI-assisted hardware design, helping to identify potential weaknesses in generated verifiers.

Technical Contribution

The SecTB-RTL framework provides a systematic approach to auditing AI-generated RTL verification plans, emphasizing consistency between generation and execution rules. Through task-level studies, the framework ensures the integrity of verification plans in the trusted production path.

Novelty

This study is the first to systematically audit the execution validity of AI-generated RTL verification plans, highlighting the importance of consistency between generation and execution rules.

Limitations

  • Generated verification plans may fail to detect security regressions even with high structural coverage.
  • Production semantic validation passed only 9 calls, indicating mismatches in generation and execution rules.

Future Work

Future work will include improving consistency between generation and execution rules and developing more effective verification plan generation methods.

AI Executive Summary

In hardware design, AI-generated RTL verification plans can satisfy provider schemas but may fail at the boundary of trusted execution. The SecTB-RTL framework audits this consistency issue across 31 tasks and 124 hardware security regressions. The study found that while most responses were accepted by providers, only a few passed production semantic validation, revealing mismatches in generation and execution rules. This research emphasizes the importance of consistency between generation and execution in verification plans, providing a methodological foundation for future AI-assisted hardware design. Although the study reveals limitations in generated verification plans, it offers directions for improving verification plan generation methods.

Deep Analysis

Background

With the advancement of AI technology, its application in chip design is shifting from code completion to generating verification artifacts. AutoBench has demonstrated that large language models can generate self-checking HDL testbenches from design descriptions, but this introduces a trust calibration problem. While generated artifacts can parse, run, and cover much of the design, they may lack the stimulus or oracle needed to detect security property violations.

Core Problem

AI-generated RTL verification plans may fail to provide measurable security evidence when evaluated under standard simulation, synthesis, and formal tools, despite satisfying provider schemas.

Innovation

The SecTB-RTL framework ensures that each generated verification plan remains unchanged in the trusted production path through task-level studies. The framework is designed to ensure that generated verification plans can detect hidden CWE-specific regressions when evaluated under standard simulation, synthesis, and formal tools.

Methodology

  • �� Use a deterministic non-AI baseline for comparison
  • �� Task-level studies ensure generated verification plans remain unchanged in the trusted production path
  • �� Evaluate generated verification plans under standard simulation, synthesis, and formal tools

Experiments

The experimental design includes 31 tasks and 124 hardware security regressions, using a deterministic non-AI baseline for comparison. In the experiments, generated verification plans are evaluated under standard simulation, synthesis, and formal tools to detect hidden CWE-specific regressions.

Results

The study found that while most responses were accepted by providers, only a few passed production semantic validation, revealing mismatches in generation and execution rules. The deterministic baseline killed 36, 75, and 78 mutants at increasing resource limits, demonstrating benchmark feasibility.

Applications

This study provides a methodological foundation for AI-assisted hardware design, helping to identify potential weaknesses in generated verifiers. It offers directions for improving verification plan generation methods.

Limitations & Outlook

Generated verification plans may fail to detect security regressions even with high structural coverage. Production semantic validation passed only 9 calls, indicating mismatches in generation and execution rules.

Plain Language Accessible to non-experts

Imagine you're cooking in a kitchen. An AI-generated RTL verification plan is like a recipe that tells you how to make a dish. You prepare all the ingredients according to the recipe and start cooking. However, despite the recipe looking perfect, you find that the final dish doesn't taste as expected. This is because some steps in the recipe didn't account for actual cooking conditions like heat and time. The SecTB-RTL framework is like a chef who checks if each step is executed correctly, ensuring the final dish meets the expected taste.

ELI14 Explained like you're 14

Imagine you're playing a game where you have to build a fortress. The AI-generated RTL verification plan is like a building guide in the game, telling you how to build the fortress step by step. You follow the guide but find that the final fortress isn't stable. This is because the guide didn't consider some important details like material choice and structural stability. The SecTB-RTL framework is like a game master who checks each step's execution, ensuring the final fortress is strong and durable.

Glossary

RTL (Register Transfer Level)

RTL is an abstraction level used to describe the behavior of digital circuits, commonly used in hardware design and verification.

In the paper, RTL is used to describe the behavior of hardware designs for verification purposes.

CWE (Common Weakness Enumeration)

CWE is a classification system for software and hardware security vulnerabilities, helping to identify and fix security issues.

In the paper, CWE is used to identify security regressions in hardware designs.

Mutation Testing

Mutation testing is a software testing technique that introduces faults to evaluate the effectiveness of test cases.

In the paper, mutation testing is used to evaluate the effectiveness of generated verification plans.

Verification Plan

A verification plan is a set of steps and tests used to verify the correctness of a hardware design.

In the paper, verification plans are generated by AI to detect security regressions in hardware designs.

Production Semantic Validation

Production semantic validation refers to verifying the effectiveness of generated verification plans in actual production environments.

In the paper, production semantic validation is used to assess whether generated verification plans meet production standards.

Open Questions Unanswered questions from this research

  • 1 How to improve consistency between generation and execution rules to enhance verification plan effectiveness?
  • 2 How to detect undetected security regressions even with high structural coverage?

Applications

Immediate Applications

Hardware Design Verification

The SecTB-RTL framework can be used to verify AI-generated RTL verification plans, ensuring their effectiveness in production environments.

Long-term Vision

AI-Assisted Hardware Design

By improving consistency between generation and execution rules, AI can more effectively assist in hardware design and verification.

Abstract

AI-generated RTL verification plans can satisfy a provider schema yet fail at the boundary to trusted execution. We present SecTB-RTL, an auditable framework covering 31 tasks and 124 authored hardware-security regressions. A deterministic non-AI baseline killed 36, 75, and 78 mutants at increasing resource limits. The first confirmatory run (C1-R2) failed before model execution because the provider rejected its response schema. After a schema-only repair made without viewing outcomes, a separately frozen follow-up run (C1-R3) completed 1,860 calls. The provider accepted 1,857 responses, but only nine passed the production semantic validator. The generation and execution rules did not match. We therefore preserve the run as an instrument-validation incident and report no prompt-effect estimate. This incident shows that provider or schema acceptance does not establish execution validity. Compilation and coverage are only diagnostics; the exact saved artifact must pass the full production path. A subsequent follow-up is excluded because it did not satisfy the preregistered evidence-completeness gate and is treated only as future work. We release the benchmark, failure-preserving contract, incident provenance, and governance controls needed to prevent infrastructure behavior from being misreported as model behavior.

cs.CR cs.AI cs.CE cs.IR cs.LG