OpenZeppelin finds data contamination in OpenAI’s EVMbench

CoinTelegraph

by Brian Quarmby

March 3, 2026

AI-Generated Deep Dive Summary

Security auditor OpenZeppelin has identified significant issues in OpenAI’s EVMbench project, a blockchain security benchmark designed to evaluate AI models' ability to detect and address smart contract vulnerabilities. During its audit, OpenZeppelin discovered methodological flaws and data contamination in the dataset used for training and testing. These findings include training data leaks and at least four invalid high-severity vulnerability classifications. EVMbench was launched in mid-February in collaboration with crypto investment firm Paradigm to assess AI models' performance in identifying, patching, and exploiting smart contract vulnerabilities. OpenZeppelin welcomed the initiative but decided to subject EVMbench to the same rigorous scrutiny it applies to major protocols like Aave, Lido, and Uniswap. In an X post on Monday, the firm highlighted its commitment to ensuring robust security standards across blockchain projects. The discovery of data contamination raises concerns about the accuracy and reliability of EVMbench as a benchmarking tool for AI models in blockchain security. OpenZeppelin’s findings emphasize the importance of thorough auditing and transparency in such projects. While OpenAI and Paradigm have not yet responded to these allegations, the issue underscores the challenges of developing reliable AI tools for cybersecurity in the cryptocurrency space. This development matters to the crypto community because it highlights potential risks in relying on AI models for critical security tasks. If EVMbench’s dataset is flawed, it could lead to inaccurate assessments of vulnerabilities, potentially compromising the security of smart contracts and decentralized systems. OpenZeppelin’s critique serves as a reminder of the need for rigorous validation processes when integrating AI into blockchain security frameworks. Overall, while EVMbench was intended to advance AI capabilities in identifying vulnerabilities, OpenZeppelin’s findings underscore the importance of addressing

Verticals

cryptoblockchain

Originally published on CoinTelegraph on 3/3/2026