What It Is
SuiAudit is an AI-driven audit tool for the Sui blockchain ecosystem, built to improve smart contract security and code quality using machine learning models. It combines AI with blockchain security auditing to help developers detect vulnerabilities and optimize Move language code. It is an emerging open-source project, and some of its infrastructure — SuiAudit Raw1 — has been released on GitHub for community contributions and reproducibility.
Sui is known for its high throughput and parallel execution, but the complexity of the Move language — object ownership, capabilities, dynamic fields — makes smart contracts prone to vulnerabilities, and so to financial losses and security risks. Those are the same features that took center stage when building an on-chain chat app at the Sui Bootcamp. Traditional auditing relies on human experts, which is costly, time-consuming, and unable to keep up with the rapid iteration of Web3 development. SuiAudit addresses that with automated, cost-effective auditing, optimized for small models that can run in resource-constrained environments and designed to run fully offline, so code need not be sent to a cloud service. The aim is to let small and medium-sized developers reach professional-level auditing rather than leaving it to large enterprises.
My Role
I worked on dataset collection and model training.
How It Works
The core product is SuiAudit Pure1, a small language model fine-tuned with extremely low resources and dedicated to smart contract auditing. It detects vulnerabilities, assesses code risks, and generates audit reports automatically.
- Pure1 is fine-tuned specifically for Sui Move rather than adapted from a general-purpose code model.
- The model is small enough to run offline in resource-constrained environments, so code is never uploaded to a cloud service.
- The Walrus protocol is integrated for decentralized storage.
- Users can upload code for real-time analysis, and POCs such as Sui Wallet Cleanup run inside the Sui ecosystem.
- Experimental code is published as Kaggle notebooks, so other developers can compare Pure1’s performance against other LLMs.
- At the 2025 Walrus Haulout Hackathon the team demonstrated how AI models can quickly generate risk assessment reports.
On the team’s own Sui-specific benchmark, Pure1 reached 31.4% accuracy, ahead of Claude Opus 4.5 (26.6%) and GPT-5.2 (17.2%).
The methodology and benchmark results are written up separately in the Pure1 research paper.
What’s Next
The plan is to expand from contract auditing to full blockchain network auditing, to bring in more AI techniques such as reinforcement learning, and to develop dedicated LLMs for handling complex ecosystems. Alongside that, the team intends to compile high-quality audit datasets, support hybrid manual-plus-automated auditing modes, and open a free online service. Possible directions include collaborating with the Sui Foundation to integrate into official toolchains, expanding to other chains such as Ethereum or Aptos, and launching paid advanced features such as real-time monitoring and custom rules. Because the work is open on GitHub and Kaggle, model improvements can come from the community, which could turn it into a crowdsourced audit platform.