OpenAI and Paradigm launch EVMbench: a security benchmark for crypto-coding agents

#What happened

• OpenAI and crypto VC firm Paradigm released EVMbench, a tool that evaluates how well AI agents can identify, patch, or exploit smart-contract vulnerabilities.
• The benchmark is built from 120 vulnerabilities drawn from 40 prior smart-contract audits, plus vulnerability scenarios from audits of Paradigm's forthcoming Tempo blockchain.
• Benchmark results: OpenAI's GPT-5.3-Codex more than doubled the effectiveness of the earlier GPT-5 at exploiting smart-contract vulnerabilities; Claude Opus 4.6 scored the highest mean result at detecting vulnerabilities, while GPT-5.3-Codex scored highest at patching and exploiting.
• OpenAI said agents "remain below full coverage" on finding and fixing bugs: they perform best in the exploit setting and are weaker on detect and patch tasks.
• The release comes days after a bug in AI-generated code cost users of crypto protocol Moonwell nearly $2.7 million; one Moonwell engineer said the code had passed an audit from security firm Halborn. DefiLlama data puts total crypto hacks and exploits in 2026 at over $108 million.

#Why it matters

• As AI agents write more production code, the industry is starting to measure not just whether agents can ship features but whether they can spot and fix the vulnerabilities they may introduce.
• EVMbench shows a clear split: agents are good at breaking contracts and weaker at securing them. OpenAI itself cautions the benchmark has a limited vulnerability sample and cannot reliably judge whether agent-found vulnerabilities are false positives.