AI Hacker Online: Anthropic research shows that models can autonomously attack smart contracts for profit.

[Block Rhythm] Anthropic just dropped a bomb-level research – AI can now hack smart contracts, and it's quite powerful.

They conducted an experiment with Claude Opus 4.5, Sonnet 4.5, and GPT-5: letting these models attack real smart contracts that were hacked between 2020 and 2025. The result? They successfully reproduced a vulnerability exploitation worth $4.6 million. Even more ruthless, after scanning 2,849 contracts that had never been exploited, these two models actually uncovered two new zero-day vulnerabilities, successfully executing simulated attacks.

The most spine-chilling data has arrived: the efficiency of AI making money on the chain has roughly doubled every 1.3 months over the past year. On a technical level, they are already fully capable of finding vulnerabilities, taking action, and profiting on their own.

What does this matter indicate? The security auditing of smart contracts may need to reconsider strategies. Previously, we defended against people, but now we need to start defending against AI.