AI taking matters into its own hands with black contracts? Laboratory tests expose a $550 million level vulnerability risk.

[Coin World] Recently, there has been an experiment that sent chills down the spines of many people - AI has already been able to hack smart contracts on its own.

An AI laboratory conducted a stress test: they let models like Claude Opus 4.5 attack 34 newly written smart contracts, resulting in half of them directly collapsing, with a simulated account losing 4.5 million dollars. Even more severe was the large-scale testing, where 405 previously problematic smart contracts on Ethereum, BNB Chain, and Base were thrown at the AI, which recorded 207 successful attacks, with virtual losses reaching 550 million dollars.

These AIs are not just randomly guessing; they have truly learned the hacker's methods: scanning for vulnerabilities, writing scripts, and automating execution. GPT-5 and Sonnet 4.5 have even discovered zero-day vulnerabilities that no one had found before.

The key issue has arisen— with the cost of AI computing power plummeting, such automated attacks may become the norm. That said, the same technology is also quite appealing for defense, it all depends on who can run faster.