Early results show significant performance differences between GPT-5.3-Codex and GPT-5, highlighting rapid model advancement.
OpenAI has unveiled EVMbench, a smart contract security benchmark developed alongside crypto investment firm Paradigm to test artificial intelligence agents on Ethereum vulnerabilities. The framework is intended to determine whether AI systems can detect, exploit and fix serious flaws in Ethereum smart contracts.
Because smart contracts are generally immutable once deployed, errors can have enduring financial consequences. OpenAI said such contracts routinely protect more than US$100 billion (AU$141 billion) in open-source crypto assets, increasing the importance of rigorous security evaluation as AI coding capabilities advance.
Measuring AI Performance
The dataset underpinning EVMbench consists of 120 curated vulnerabilities drawn from 40 professional audits, with most sourced from open audit competitions including Code4rena. Additional scenarios stem from security auditing work for Tempo, a purpose-built Layer-1 blockchain designed to support high-throughput, low-cost stablecoin payments.
AI agents are assessed across three categories: detecting known vulnerabilities, patching contracts without compromising intended functionality, and executing exploit attempts within a controlled blockchain environment. Exploit tasks are graded using deterministic transaction replay and on-chain checks.
In benchmark results, GPT-5.3-Codex achieved 72.2% in exploit mode, while GPT-5 recorded 31.9%, despite being released just over six months earlier. OpenAI said the objective is to create a clear standard for evaluating AI systems in blockchain security as decentralised finance continues to grow.
Read more on Crypto News Australia
