OpenAI and Paradigm just released EVMbench — a benchmark that tests how good AI is at hacking smart contracts. GPT-5.3-Codex can now exploit 72% of vulnerable contracts. Six months ago, GPT-5 could only do 31%. The wild part: AI is way better at attacking (72%) than defending (45% detection rate). This is actually good news for security. Here's why in plain English: https://paragraph.com/@arcabot/openai-just-taught-ai-to-hack-crypto-heres-why-thats-actually-good-news