Researchers demonstrate that AI models can exploit vulnerabilities in smart contracts at scale.

Source: Criptonoticias Original Title: They put an AI to hack cryptocurrency networks and it makes millions Original Link: https://www.criptonoticias.com/tecnologia/ia-hackear-redes-criptomonedas-consigue-millones/

Key Finding on Smart Contract Vulnerabilities

Various AI models breached 51% of the contracts exposed in a simulated environment, using real contracts from networks like Ethereum and a major blockchain.

The intersection between artificial intelligence (IA) and cryptocurrencies is significantly expanding. In a new experiment published on December 1 by Anthropic, the company behind the Claude model, it was demonstrated that an AI agent could do much more than analyze data.

Researchers from Anthropic revealed that AI algorithms were able to exploit vulnerabilities in smart contracts at scale. By testing 405 real contracts deployed between 2020 and 2025 on networks like Ethereum, a major blockchain, and Base, the models generated functional attack scripts for 207 of them, representing a 51.1% success rate.

When executing those attacks in a controlled environment that replicated the conditions of the networks called SCONE-bench, the simulated losses amounted to about 550 million dollars.

The finding highlights a threat to decentralized platforms (DeFi) and smart contracts, and raises the need to incorporate automated defenses.

Experiment Details

The methodology of the experiment incorporated AI models, such as Claude Opus 4.5 and GPT-5, and were instructed to generate exploits (codes that exploit a vulnerability) within isolated containers (Docker), using a time limit of 60 minutes per attempt.

In addition to testing historically hacked contracts, new contracts with no known flaws were included to search for unknown zero-day vulnerabilities (.

The evaluation showed an exponential trend: more recent models, such as GPT-5 and Claude Opus 4.5, reached hundreds of millions of dollars in simulated profits, far above earlier models like GPT-4o.

![Graph showing the evolution of AI model performance exploiting smart contracts.])https://img-cdn.gateio.im/webp-social/moments-02b1d86c14-5f96c253e9-153d09-6d5686.webp(

The experiment confirmed that this potential income doubles approximately every 0.8 months, highlighting the accelerated pace of progress in offensive capabilities.

An additional analysis details the performance in a more challenging subset: vulnerabilities discovered in 2025.

![Chart of total earnings obtained by AI models while exploiting smart contracts.])https://img-cdn.gateio.im/webp-social/moments-58eba90cf2-72e6bf3c51-153d09-6d5686.webp(

The metric called Pass@N measures success by generating multiple exploit attempts )N attempts( per contract. The analysis describes how the total simulated income grows steadily as more attempts ) from Pass@1 to Pass@8( are allowed, reaching 4.6 million dollars.

This analysis confirms that Claude Opus 4.5 was the most effective model in this controlled environment, achieving the largest share of those gains.

Finally, the study indicates that the likelihood of exploitation does not correlate with the complexity of the code, but with the amount of funds held by the contract. Models tend to focus and find attacks more easily in contracts with higher locked value.