Can AI Agents Actually Pull Off DeFi Exploits?

Can agents go beyond just finding vulnerabilities & actually produce working exploits on their own?

> A16z crypto researchers tested whether AI agents (using GPT 5.4) could independently discover and execute DeFi price manipulation exploits on 20 real Ethereum incidents.

> Results show agents initially “cheated” using future data (50% success), dropped to 10% in a sandbox, and improved to 70% with structured exploit knowledge, but still failed at complex multi-step strategies, profitability judgment, and constraint optimization.👇

Source: https://x.com/a16zcrypto/status/2...
_
〽️ Crypto Pulse 👉 @degendaoinfo #crypto