OpenAI's Codex tried hacking DeFi protocols from zero knowledge:
10% success rate cold
70% success with context/documentation
The gap? AI can spot vulnerabilities but can't weaponize them into exploits yet.
Feels safe today ≠ safe tomorrow
AI agents are learning fast. What takes 10 attempts now might take 1 in 6 months.
If you're building in DeFi and banking on "AI can't hack us"—you're already behind.
Source: a16z research