Introducing SuiAudit Pure1: Revolutionizing Sui Smart Contract Security with AI

I’m excited to share a major milestone for our team at SuiAudit Lab: the publication of our research paper, “Can Small Models Beat Frontier LLMs? Introducing SuiAudit Pure1: Specialized Auditing via Extreme Low-Resource Fine-Tuning,” now live on TechRxiv (DOI: 10.36227/techrxiv.177204934.48017152/v1). As a key contributor alongside Chia-Lin Chen and Fang-Yu Lien, this work represents our passion for bridging AI and blockchain security, specifically tailored for the Sui ecosystem.

The Core Innovation

SuiAudit Pure1 is a 7B-parameter AI model designed from the ground up to audit Sui Move smart contracts with unprecedented accuracy and efficiency. Unlike general-purpose large language models (LLMs) plagued by “Solidity Bias”—where they mistakenly apply Ethereum vulnerabilities to Sui’s unique object-centric architecture—Pure1 excels through a streamlined fine-tuning pipeline: Supervised Fine-Tuning (SFT) for domain mastery, followed by Direct Preference Optimization (DPO) enhanced with Chain-of-Thought reasoning.

Our secret sauce? A meticulously crafted dataset blending positive vulnerability samples (categorized A-F) with adversarial “Hard Negative” examples—safe code patterns engineered to trip up broader LLMs. This allows Pure1 to run fully offline, preserving developer privacy while delivering real-time insights via VS Code and Chrome extensions.

Standout Results and Impact

The results speak for themselves: On our rigorous Sui-specific benchmark, Pure1 achieved a 31.4% accuracy rate, outperforming industry giants like Claude Opus 4.5 (26.6%) and GPT-5.2 (17.2%). We’ve already deployed it in production, securing over $50M in assets, detecting 500+ vulnerabilities, and earning accolades like 2nd Place at the 2025 Sui Hackathon Mandarin and 3rd at the Sui Taipei Dev Hackathon.

This isn’t just a model—it’s a game-changer for Sui developers, enabling faster, safer deployments in Asia’s booming Web3 scene. By open-sourcing our core infrastructure on GitHub and sharing weights on Kaggle, we’re fostering a more secure blockchain future.

Reflections on the Journey

Publishing this paper feels like a validation of our low-resource approach: proving that specialized small models can punch above their weight against frontier LLMs. It’s a reminder that innovation thrives at the intersection of AI and blockchain. Stay tuned as we continue scaling SuiAudit—your feedback and collaborations are welcome!

If you’re building on Sui or interested in AI-driven security, check out the full paper and try our live tool at suiaudit.wal.app. Let’s make Web3 safer together.

論文連結在這裡
 官方網站

SuiAudit Pure1 發表：以 AI 革新 Sui 智慧合約安全

我很高興分享 SuiAudit Lab 團隊的重要里程碑：研究論文〈Can Small Models Beat Frontier LLMs? Introducing SuiAudit Pure1: Specialized Auditing via Extreme Low-Resource Fine-Tuning〉已在 TechRxiv 發表（DOI: 10.36227/techrxiv.177204934.48017152/v1）。我和陳家霖、連方妤共同參與這份研究；它展現我們將 AI 與區塊鏈安全結合、並專為 Sui 生態系打造解決方案的熱情。

核心創新

SuiAudit Pure1 是從零開始設計、擁有 70 億參數的 AI 模型，目標是在 Sui Move 智慧合約稽核上兼具前所未有的準確度與效率。不同於通用大型語言模型（LLM）常受「Solidity 偏誤」影響，錯把 Ethereum 的漏洞套用到 Sui 獨特、以物件為核心的架構，Pure1 採用精簡的微調流程：先以監督式微調（SFT）建立領域能力，再使用融入思維鏈推理的直接偏好最佳化（DPO）。

關鍵在於精心製作的資料集：結合 A 至 F 類的正向漏洞樣本，以及刻意設計、用來讓通用 LLM 誤判的安全程式模式「Hard Negative」對抗樣本。這讓 Pure1 能完全離線運作，保護開發者隱私，並透過 VS Code 與 Chrome 擴充功能提供即時洞察。

突出的成果與影響

結果相當明確：在嚴謹的 Sui 專屬基準測試中，Pure1 達到 31.4% 準確率，優於 Claude Opus 4.5 的 26.6% 與 GPT-5.2 的 17.2%。我們已將它部署至正式環境，保護超過 5,000 萬美元資產、偵測逾 500 個漏洞，並獲得 2025 Sui Hackathon Mandarin 第二名與 Sui Taipei Dev Hackathon 第三名等肯定。

這不只是一個模型，而是 Sui 開發者的關鍵助力，能讓亞洲快速成長的 Web3 場景更快、更安全地部署應用。透過在 GitHub 開源核心基礎設施、於 Kaggle 分享模型權重，我們正共同打造更安全的區塊鏈未來。

歷程反思

這篇論文的發表驗證了低資源方法的價值：專用小型模型也能在前沿 LLM 面前展現超越規模的實力。它提醒我們，創新往往發生在 AI 與區塊鏈的交會處。SuiAudit 還會持續擴展，歡迎提供回饋與合作。

若你正在 Sui 上開發，或對 AI 驅動安全有興趣，歡迎閱讀完整論文，並到 suiaudit.wal.app 試用線上工具。讓我們一起讓 Web3 更安全。

Introducing SuiAudit Pure1：Revolutionizing Sui Smart Contract Security with AISuiAudit Pure1 發表：以 AI 革新 Sui 智慧合約安全