Morning Minute: OpenAI and Paradigm Turn Focus to Smart Contracts

Morning Minute is simply a regular newsletter written by Tyler Warner. The investigation and opinions expressed are his ain and bash not needfully bespeak those of Decrypt. Subscribe to the Morning Minute on Substack.

GM!

Today’s apical news:

• Crypto majors autumn 1-2% amidst Iran escalations; BTC astatine $66.6k

• Coinbase’s Base announced it’s dropping the OP stack for its ain unified lawsuit (OP -23%)

• Brian Armstrong called the looming quantum hazard a “solvable issue” for crypto

• Hyperliquid launched its Hyperliquid Policy Center with $29M for DC advocacy

• MegaETH introduced its World Markets arsenic archetypal DEX to unify spot trading, lending and perps each onchain with transverse margin

Smart contracts have mislaid billions to exploits.

Now the world’s astir almighty AI laboratory and crypto’s astir respected probe steadfast are gathering the tools to extremity that.

OpenAI and Paradigm jointly released EVMbench yesterday, an unfastened benchmarking model that evaluates AI agents crossed 3 modes:

• detecting vulnerabilities successful Ethereum astute contracts

• patching them

• actively exploiting them.

The benchmark draws connected 120 curated high-severity vulnerabilities pulled from 40 real-world audits, mostly sourced from Code4rena competitory audit contests. It besides includes scenarios from the information audit of Stripe’s Tempo, the payments-focused L1 built with input from Visa, Shopify, and OpenAI.

When Paradigm started this project, apical models could exploit less than 20% of captious bugs. That fig is present supra 70%.

OpenAI besides expanded the backstage beta of Aardvark, its dedicated information probe agent, and committed $10M successful API credits done its Cybersecurity Grant Program to enactment antiaircraft crypto research.

Can AI Agents Boost Ethereum Security? OpenAI and Paradigm Created a Testing Ground

Paradigm, from their EVMbench release: “It’s present wide to america that a increasing information of audits successful the aboriginal volition beryllium done by agents. Hopefully this benchmark, harness, and cause service some arsenic a preview and an accelerant towards that future.”

OpenAI noted that measuring AI show successful “economically applicable environments is captious arsenic models go almighty tools for some attackers and defenders.”

Smart declaration exploits person drained implicit $5B from DeFi successful the past 2 years alone.

EVMbench is OpenAI and Paradigm’s determination to halt the drain.

The irony?

The aforesaid capableness that makes EVMbench almighty (AI that tin find and exploit bugs astatine 72% accuracy) is besides the nonstop menace exemplary the instrumentality defends against. In the incorrect hands, a benchmark this precise becomes a hacking playbook.