AI models can now hack blockchain contracts: Study
AI is getting surprisingly good at breaking into blockchain smart contracts, according to a new study from Anthropic.
Their tests showed that models like Claude Opus 4.5 and GPT-5 could pull off past attacks on Ethereum-style blockchains, with Opus 4.5 alone grabbing about $4.5 million in simulated funds from a specific dataset.
Even more striking—the success rate of these hacks is doubling every 1.3 months, while the cost to run them has dropped by 70% in just half a year.
AI: Both a threat and a shield for blockchain
Here's the twist: while AI can easily spot and exploit known issues, it struggles with finding brand-new vulnerabilities—mostly catching only basic bugs in untouched contracts.
On the flip side, developers are using AI to scan code for flaws before launch, making it both a hacker's tool and a defender's ally in the ongoing tug-of-war over blockchain security.