EVMbench is built from 120 curated vulnerabilities across 40 audits and incorporates data from security processes used in Stripe’s Tempo project to simulate real-world scenarios.

There is rising interest in agentic AI for crypto, with predictions that AI agents could influence stablecoin payments and become integrated into crypto-native workflows for transactions.

The piece appears on a sponsor-heavy Techmeme-style page, but the EVMbench item is the substantive narrative within that sponsorship context.

The Block notes AI-agent capabilities in security discussions, linking to related research and industry activity around AI-driven vulnerability identification and risk.

Initial reporting is attributed to Cointelegraph and hokanews, with additional technical documentation expected for methodology and benchmarks.

OpenAI and Paradigm have launched EVMbench, a benchmark designed to test AI agents’ ability to detect, patch, and exploit high-severity smart contract vulnerabilities within a sandboxed blockchain environment.

Looking ahead, AI-driven security tooling could reduce development costs and time-to-market for secure DApps, while ongoing research probes AI capabilities and limitations in code analysis.

In performance, Anthropic’s Claude Opus 4.6 led results with an average detect award of about $37,824, followed by OpenAI’s OC-GPT-5.2 and Google’s Gemini 3 Pro.

The initiative signals growing AI involvement in DeFi security through a collaboration between an AI research organization and a crypto-focused investment firm.

CoinDesk frames the story within its weekly tech-crypto wrap, noting related industry developments such as Ethereum Foundation leadership changes and XRP Ledger progress.

The article acknowledges AI-assisted generation and references related policy and coverage, with AI workflows contributing to the piece itself.

AI workflows assisted in producing the article, reflecting the benchmark’s broader context of AI’s role in security discourse.