- OpenAI and Paradigm have launched EVMbench, a new tool to test AI agents on real Ethereum smart contract vulnerabilities.
- The benchmark evaluates three critical modes: detecting, patching, and exploiting high-severity security flaws.
- In exploit tests, the latest GPT-5.3-Codex scored 72.2%, more than doubling the performance of its predecessor.
- The tool uses 120 curated vulnerabilities from 40 audits, with many sourced from open competitions like Code4rena and Tempo.
On February 18, 2026, OpenAI, in collaboration with investment firm Paradigm, introduced EVMbench, a groundbreaking tool designed to evaluate AI agents on Ethereum smart contract security. The Ethereum Virtual Machine test suite focuses on high-severity vulnerabilities as smart contract deployment reaches record highs. Consequently, the benchmark draws on 120 real-world vulnerabilities curated from 40 audits, including scenarios from the security process for Tempo.
Stripe launched the public testnet for its purpose-built blockchain, Tempo, in December with input from VISA, Shopify, and OpenAI. The goal is to ground testing in economically meaningful code, especially as AI-driven stablecoin payments expand. EVMbench tests agents across three distinct modes: detect, patch, and exploit. In the exploit phase, agents attempt end-to-end fund-draining attacks within a sandboxed environment.
Testing revealed that GPT-5.3-Codex achieved a 72.2% success rate in exploit mode, according to an OpenAI blog post. This significantly outperformed the 31.9% score of GPT-5, released six months earlier. However, performance was weaker in the detect and patch tasks, where agents sometimes failed to audit exhaustively.
Researchers cautioned that EVMbench does not fully capture real-world security complexity. Meanwhile, the weekly number of smart contracts deployed on Ethereum reached 669,500, according to Token Terminal. Measuring AI performance in such environments is now critical as models become powerful tools for both attackers and defenders.
✅ Follow BITNEWSBOT on Telegram, Facebook, LinkedIn, X.com, and Google News for instant updates.
