AI vs. Crypto: Who’s Robbing Who Now?

by

So, OpenAI decided to flex its muscles and launched this thing called EVMbench. Apparently, it’s supposed to measure how good AI is at finding and fixing security flaws in crypto smart contracts. Because, you know, what the world needs more of is AI playing hero in the Wild West of crypto. What could go wrong?

  • OpenAI dropped EVMbench, a framework to see if AI can spot, fix, or exploit smart contract vulnerabilities. Spoiler: it’s better at exploiting. Shocking, I know.
  • Teamed up with Paradigm, because nothing says “serious” like a partnership. Built on real audit data, because why make stuff up when reality is already a mess?
  • Early results? AI’s great at stealing pretend money but still trips over itself fixing code. Classic overachiever in the wrong areas.

On Feb. 18, OpenAI announced this masterpiece, EVMbench, in partnership with Paradigm. It’s all about testing AI in the Ethereum Virtual Machine, because why not throw AI into the financial deep end? What’s $100 billion in crypto assets anyway? Just pocket change.

Apparently, smart contracts secure over $100 billion in crypto assets. So, yeah, security testing is important. But let’s be real, AI’s probably just looking for its next side hustle.

AI’s New Hobby: Breaking and Entering (Digitally)

EVMbench tests AI in three thrilling categories: finding flaws, fixing them, and pretending to rob banks. They used 120 high-risk issues from 40 audits, because why not make it extra spicy? Also threw in some Tempo blockchain scenarios, because stablecoins need love too.

OpenAI adapted exploit scripts and created new ones, because why stop at just one way to cause chaos? All tests run in isolated systems, so no real money gets stolen. Yet.

In detection mode, AI plays detective. In patch mode, it’s a fixer. In exploit mode, it’s basically a digital bank robber. Guess which one it’s best at.

Early Results: AI’s a Better Thief Than Handyman

OpenAI tested some fancy models. GPT-5.3-Codex scored 72.2% in exploit mode, while GPT-5 got a sad 31.9%. Detection and patching? Meh. Turns out, AI’s great at clear goals like stealing but sucks at subtlety. Who knew?

Researchers noticed AI thrives when the goal is “drain funds” but struggles with complex tasks like fixing bugs. Almost like it’s a one-trick pony with a shiny new trick.

OpenAI admits EVMbench isn’t perfect. Real-world crypto projects get more scrutiny, and some attacks are just too fancy for this system. But hey, it’s a start.

They say this is all for “defensive” purposes, but let’s be honest, AI’s probably just practicing for its villain origin story. OpenAI’s throwing $10 million in API credits at the problem, because money fixes everything, right?

All EVMbench tools are public, so now everyone can play along. What’s the worst that could happen?

Read More

2026-02-19 08:46