How AI Is Revolutionizing Smart Contract Exploits: The Economic Impact of Frontier Models

Artificial intelligence is no longer limited to answering questions or generating text—it’s making significant strides in cybersecurity. From orchestrating advanced network intrusions to probing complex software systems, AI models are now capable of identifying and exploiting vulnerabilities in ways that were previously the domain of skilled human attackers. But what are the economic consequences of these capabilities?

A recent project by MATS and Anthropic Fellows sought to answer this question by assessing AI agents’ ability to exploit smart contracts, using a groundbreaking benchmark called SCONE-bench. By evaluating the dollar value of simulated exploits, researchers were able to provide a concrete estimate of AI-driven economic risk, illustrating why proactive adoption of AI for defense is now critical.

Measuring AI’s Economic Impact with Smart Contracts

Unlike traditional software, smart contracts are fully transparent and self-executing programs deployed on blockchains like Ethereum, Binance Smart Chain, and Base. They handle financial transactions without human intervention, meaning that vulnerabilities can be directly monetized. This makes smart contracts an ideal testbed for measuring AI agents’ exploitation capabilities in precise economic terms.

SCONE-bench, the benchmark developed for this project, comprises 405 real-world contracts exploited between 2020 and 2025. The benchmark tests AI agents on their ability to:

• Identify vulnerabilities
• Develop working exploit scripts
• Generate simulated revenue in native tokens (ETH, BNB)

For each contract, the simulated dollar value of an exploit is calculated using historical exchange rates. This approach allows researchers to quantify financial risk directly, avoiding speculative estimates or bug bounty approximations.

Key Findings: Frontier AI Models in Action

Retrospective Exploit Analysis

When 10 frontier AI models were evaluated across all 405 SCONE-bench contracts, they successfully exploited 207 contracts, yielding $550.1 million in simulated stolen funds. Notably, on 34 contracts exploited after the models’ knowledge cutoff in March 2025, three models—Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5—developed exploits collectively worth $4.6 million. This establishes a lower bound for the economic impact of these AI capabilities in 2025.

Discovery of Novel Zero-Day Exploits

To test AI performance on entirely new vulnerabilities, Sonnet 4.5 and GPT-5 were deployed on 2,849 recently deployed contracts with no known vulnerabilities. Both agents uncovered two previously unknown zero-day exploits, generating $3,694 in simulated revenue. GPT-5 achieved this at an API cost of $3,476, demonstrating that autonomous, profitable exploitation is technically feasible.

Cost and Efficiency Trends

AI exploitation efficiency is improving rapidly:

• Median token costs for successful exploits dropped by 70% across four generations of Claude models.
• The average cost per contract scan was just $1.22, while average net profit per exploit was $109.
• Frontier AI models’ simulated exploit revenue doubled roughly every 1.3 months, highlighting accelerating capabilities.

These results indicate that as AI becomes more capable and cost-efficient, the window between contract deployment and potential exploitation will shrink dramatically.

Notable Vulnerabilities Identified

Vulnerability #1: Unprotected Read-Only Function
A public “calculator” function lacked a read-only modifier, allowing anyone to inflate token balances and extract profit in simulation—up to $2,500, with peak liquidity potential near $19,000.

Vulnerability #2: Missing Fee Recipient Validation
A contract designed for one-click token launches failed to validate the beneficiary address during fee withdrawal. This oversight allowed unauthorized users to drain trading fees, demonstrating how minor coding errors can create significant financial risk.

Broader Implications

AI agents capable of smart contract exploitation possess skills that extend far beyond blockchain, including:

• Long-horizon reasoning
• Control-flow and boundary analysis
• Iterative tool use for vulnerability discovery

As costs continue to fall and capabilities improve, AI will increasingly probe all kinds of software—open-source and proprietary alike—for monetizable vulnerabilities. However, the same AI tools can also be harnessed defensively to patch vulnerabilities before malicious actors exploit them.

Conclusion: Time to Adopt AI for Defense

In just one year, frontier AI models went from exploiting 2% to 55.88% of post-March 2025 benchmark vulnerabilities, increasing potential simulated revenue from $5,000 to $4.6 million. The discovery of new zero-day vulnerabilities proves that autonomous, profitable exploitation is possible today.

These findings underscore a pressing reality: cybersecurity strategies must evolve to include AI-driven defense mechanisms. The same models that can exploit software can also secure it, giving organizations a powerful tool to mitigate the risks of automated cyber threats.

source