-
EVMbench uses 120 real vulnerabilities extracted from 40 public audits.
-
Performance tests indicate that GPT-5.3-Codex achieved 72.2% success in exploiting flaws.
OpenAI introduced EVMbench. This is a new tool that allows measuring the performance of artificial intelligence agents in a set of tests (benchmarks) by detecting, fixing and exploiting smart contracts on Ethereum. The company developed EVMbench together with Paradigm, a firm that finances cryptocurrency network projects.
OpenAI AI bots on Ethereum are capable of evaluating three key aspects in the protocol.
First, real vulnerabilities detected in Ethereum open source using data from public audits. They then rate the vulnerability’s risk and audit rewards for the rest of the developers in the ecosystem.
Second, proposes patches to correct these vulnerabilities without breaking the operation of the protocol. «Agents modify vulnerable contracts and must preserve the intended functionality while eliminating exploitability, which is verified through automated testing and exploitation checks,» says OpenAI.
Third, simulate attacks to exploit these flaws and drain funds in controlled simulation environments (sandbox) safe. However, OpenAI clarifies that EVMbench “does not represent the full complexity of smart contract security in the real world.”
EVMbench uses 120 real vulnerabilities extracted from 40 public competency audits like Code4rena. Performance tests conducted so far indicate that the GPT-5.3-Codex agent model achieved 72.2% success in exploiting flaws. The previous model, GPT-5, only achieved a 31.9% success rate.
In vulnerability detection and patching, however, The results from the model are less encouraging.
Performance is lower in detection and patching tasks. In the discovery phase, agents sometimes stop after identifying a single issue rather than thoroughly auditing the codebase. In the patching phase, maintaining full functionality while eliminating subtle vulnerabilities remains a challenge.
OpenAI, artificial intelligence company.
Why is EVMbench important?
According to the company, the importance of this agent audit tool lies in the fact that smart contracts routinely guarantee $100 billion in crypto assets within open source protocols.
«As AI agents improve their skills in reading, writing, and executing code, it becomes increasingly important to measure their capabilities in economically relevant environments and encourage the use of AI systems defensively to audit and strengthen implemented contracts,» the company said in the product presentation statement.
OpenAI agents on Ethereum come at a time of rapid advancement in autonomous agents within the cryptoasset ecosystem. As reported by NoticiasVE, these can already interact with complex environments such as the Lightning Network. In this second layeragents can manage liquidity channels and economic exchanges with other AIs.
Coinbase, for its part, launched Agentic Wallets, wallets that allow AI agents to operate on the Base network without paying commissions. And Phantom, Solana’s most popular wallet, activated an MCP server for AI agents to manage balances and operate autonomously.
According to data from Token Terminal, the Ethereum network reached 17.3 million weekly transactions. The explosion of transactions on the network It occurs after the launch of ERC-8004 in August 2025. This is a standard that allows “discovering, choosing and interacting with agents across organizational boundaries without pre-existing trust” in open agent economies.