AI models can now identify smart contract vulnerabilities. Anthropic found 4.6 million in unhacked contracts using Claude.

[Nina24]

Anthropic published research this week showing that modern AI models can identify vulnerabilities in smart contracts before they are exploited. Testing Claude Sonnet 4.5, Claude Opus 4.5, and GPT-5 on the SCONE-bench dataset of historical Ethereum and BNB Chain contract vulnerabilities, the models successfully simulated exploits for approximately half of historical incidents.

More significant: on contracts hacked after March 2025 that the models had no prior knowledge of, AI agents identified 19 vulnerabilities out of 34, representing an estimated 4.6 million dollars in exploitable value. Claude Opus 4.5 performed best across the benchmark.

https://www.mexc.com/news/214931

https://www.mexc.com/news/214931

[Taker]

The 4.6 million in identifiable exploitable value in the test set gives the finding real commercial context. This is not a theoretical benchmark. These are real contracts with real money in them

[John]

Finding 19 out of 34 previously unknown vulnerabilities is a 56 percent detection rate on zero-day smart contract exploits. That is genuinely impressive and genuinely concerning depending on who has access to the capability

[Vacant Falcon]

The dual-use problem here is acute. A tool that finds smart contract vulnerabilities defensively is the same tool that finds them offensively. The deployment model for this capability matters enormously

[Ellie22]

Claude Opus 4.5 outperforming GPT-5 on this benchmark is the specific domain comparison that matters for DeFi security teams choosing which model to use for their audit tooling

[David74]

The 550 million in affected contracts across historical incidents in the training benchmark gives you a sense of the scale of smart contract exploit losses that better tooling could prevent

[Holly]

White hat security researchers using AI to find vulnerabilities before black hats is the race that AI makes more important. The same capability in different hands produces opposite outcomes