- OpenAI and Paradigm released EVMbench, a new tool to test AI agents on smart contract security tasks.
- Research shows AI agents are significantly better at exploiting smart contract flaws than finding or fixing them, with GPT-5.3-Codex excelling.
- The tool’s release follows a recent incident where an AI-generated bug cost Moonwell users nearly $2.7 million.
OpenAI and crypto venture firm Paradigm launched a new benchmarking tool on Wednesday that rigorously evaluates how AI agents handle smart contract security vulnerabilities. This release arrives just days after a costly bug in AI-generated code led to significant user losses.
The tool, called EVMbench, is built from 120 vulnerabilities identified in over 40 prior audits. Consequently, it provides a standardized way to measure AI performance on detection, patching, and exploitation tasks.
Results from the tool reveal a stark capability gap among current AI models. OpenAI’s latest model, GPT-5.3-Codex, more than doubled its predecessor’s effectiveness at exploiting flaws to drain funds.
However, its success in finding and fixing vulnerabilities “remain below full coverage,” according to the company’s news release. The agents sometimes stop after finding one issue or struggle to maintain functionality while patching.
In benchmark comparisons, Anthropic’s Claude Opus 4.6 scored highest for detecting vulnerabilities. Meanwhile, GPT-5.3-Codex achieved top results in both patching and exploiting smart contracts.
OpenAI cautioned that EVMbench has limitations due to its finite sample of vulnerabilities. The tool also cannot reliably determine if agent-found vulnerabilities are false positives.
Testing such tools is critical as smart contract hacks continue to plague the industry. According to data, protocols have suffered over $108 million in exploits so far in 2026.
✅ Follow BITNEWSBOT on Telegram, Facebook, LinkedIn, X.com, and Google News for instant updates.
