---
type: "Evidence Item"
title: "Introducing EVMbench"
description: "OpenAI and Paradigm introduce EVMbench, a benchmark evaluating AI agents’ ability to detect, patch, and exploit high-severity smart contract vulnerabilities."
resource: "https://openai.com/index/introducing-evmbench"
tags: ["appendix-iii", "benchmark", "openai"]
timestamp: "2026-02-18"
category: "benchmark"
publisher: "OpenAI"
cope_score: 86
confidence: 0.9
---

# Introducing EVMbench

# Claim

OpenAI and Paradigm introduce EVMbench, a benchmark evaluating AI agents’ ability to detect, patch, and exploit high-severity smart contract vulnerabilities.

# Relevance

Appendix III, section one: model and benchmark capability evidence

# Oracle Verdict

OpenAI is describing a frontier or production capability that pushes directly on the thesis. The important signal is not the marketing language; it is the widening set of tasks now being routed through model-driven execution rather than ordinary software or headcount.

# Metadata

* Publisher: OpenAI
* Category: benchmark
* Sector: General AI capability
* Capability: Frontier model release and benchmark movement
* Cope score: 86
* Confidence: 0.9

# Related Concepts

* [Live evidence index](index.md)
* [Thesis](../thesis.md)

# Citations

[1] [Introducing EVMbench](https://openai.com/index/introducing-evmbench)