Developed in collaboration with Paradigm, EVMbench evaluates AI agents’ ability to detect, patch, and exploit smart contract vulnerabilities.
EVMbench, a benchmarking tool, is set to enhance blockchain security by measuring the capabilities of AI agents in detecting, patching, and exploiting vulnerabilities in smart contracts. This new tool underscores the growing role of artificial intelligence in enhancing the security of decentralized finance (DeFi) ecosystems.
EVMbench employs historical vulnerabilities and a Rust-based harness to evaluate AI performance. At the forefront is GPT-5.3-Codex, an AI model developed by OpenAI, which achieved a score of 72.2% in exploit-mode evaluations.
EVMbench’s evaluation is comprehensive, utilizing 120 curated vulnerabilities from over 40 audits. These include scenarios provided by Tempo L1, which focuses on payment-oriented evaluations.
The tool also benefits from Paradigm’s expertise, which provides domain knowledge and quality control. This collaboration ensures the accuracy and reliability of EVMbench’s evaluations.
This article was generated with the assistance of AI workflows.
