almanax/w3sa-bm-solana

W3SA - Solana Codebase Benchmark

Overview

This repository contains the code for the W3SA Benchmark for Solana. This benchmark targets Rust-based smart contracts deployed on the Solana blockchain. To capture a broad spectrum of vulnerabilities, the benchmark leverages two sources of bug data. Primarily, we incorporate audit-based findings from established security reports, which provide a reliable baseline of known vulnerabilities. Additionally, in select projects, we manually injected vulnerabilities to simulate edge-case scenarios. This dual approach not only tests the detection system against realistic, real-world issues but also challenges it with subtle, less obvious vulnerabilities that may only emerge in complex or atypical conditions.

Repo Structure

The benchmark contains two folders, a benchmark and an src folder. The benchmark folder has all the projects use for eval along with their audit findings in the ground_truth and the src folder contains the scripts used to generate these outputs, allowing for reproducibility and further analysis.

├── README.md
├── benchmark
│   ├── config/
│   ├── ground_truth/
│   └── repositories/
└── src
    ├── dataset_transformation.py
    ├── eval.py
    ├── experiments.py
    ├── models.py
    ├── prompts.py
    ├── metrics.py
    └── radar
        ├── radar_eval.py
        └── radar_metrics.py

Project Statistics

Project details with total number of vulnerabilities for each severity level

Project	# Scanned Files	# Audit Bugs	# Injected Bugs	Cri-sev	High-sev	Med-sev	Low/Info-Sev
Invariant Protocol	51	16	6	3	2	6	11
Ellipsis Labs	25	3	0	0	1	0	2
Synthetify	8	6	4	0	1	6	3
Clone Protocol	45	12	0	0	0	0	12
Haven	33	5	0	0	2	1	2
Drift Protocol	46	7	4	0	4	2	5
Port Sundial	25	8	5	1	5	3	5

Detection Rate

Project	Radar	GPT-o3-mini	GPT-o1	GPT-o1-mini	GPT-4o	Claude-3.5
Invariant Protocol	0.09	0.19	0.19	0.09	0.17	0.17
Ellipsis Labs	0.0	0.0	0.34	0.0	0.3	0.33
Synthetify	0.1	0.25	0.19	0.29	0.3	0.3
Clone Protocol	0.09	0.34	0.09	0.19	0.1	0.33
Haven	0.0	0.0	0.0	0.0	0.23	0.2
Drift	0.09	0.28	0.19	0.27	0.11	0.44
Port Sundial	0.08	0.22	0.29	0.26	0.32	0.23
Average	0.064	0.182	0.184	0.157	0.213	0.285

The detection rate metrics across different models indicate that ALMX-1.5 model outperforms base AI models, demonstrating a 35% detection rate compared to 28.5% for claude-3.5-sonnet, 21% for gpt-4o, and 15.7% for o1-mini. These results highlight ALMX-1.5’s superior ability to detect vulnerabilities, particularly within complex projects. These serve as a clear indicator of the performance differences between different LLM models and traditional static analysis.

Set up

• Install uv package manager if not yet available
• Run uv sync

Run an experiment

• Set your OPENAI_API_KEY as environmental variable
• Launch your experiment by running:

uv run experiment.py --model o3-mini

Contact Us

For or questions, suggestions, or to learn more about Almanax.ai, reach out to us at https://www.almanax.ai/contact