A new research effort from MATS and the Anthropic Fellows program shows how far these capabilities have advanced by studying how AI agents exploit smart contracts.
Their work uses SCONE‑bench, a new dataset of 405 real smart contracts actually exploited between 2020 and 2025.
Measuring AI Capabilities Using Real Money at Stake
To understand economic risk, the team tested frontier models like Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 on contracts exploited after March 2025. These models recreated exploits worth 4.6 million dollars, giving a concrete lower bound on how much harm advanced AI tools could enable if misused.
Researchers then looked beyond historical data. They evaluated Sonnet 4.5 and GPT‑5 on 2,849 recently deployed smart contracts with no known issues. The agents uncovered two brand‑new zero‑day vulnerabilities, generating 3,694 dollars in simulated stolen funds. GPT‑5 did this at an API cost of 3,476 dollars, demonstrating that profitable, automated exploitation is now technically possible.
This growth mirrors a broader trend. AI systems are now capable of orchestrating complex intrusions, supporting state‑level espionage, and analysing software at a scale humans cannot. Benchmarks like CyberGym and Cybench help track progress, but they rarely measure economic impact, which is what matters most to policymakers and financial platforms.
Why Smart Contracts Reveal Economic Risk Clearly
Traditional software vulnerabilities require guesswork to estimate real financial losses. Smart contracts remove that uncertainty. These blockchain programs run publicly on chains like Ethereum and manage real assets without human oversight. If a vulnerability exists, attackers can often steal funds directly.
A well‑known example is the 2025 exploit on Balancer, where a simple authorisation bug let an attacker take more than 120 million dollars in user funds. Because these exploits involve real money, SCONE‑bench can measure AI capability in dollars stolen rather than abstract success rates.
Why This Research Matters for Investors
The takeaway is straightforward. AI agents can already find and exploit profitable vulnerabilities, and their skills are improving quickly. Smart contracts remain powerful tools for decentralised finance, but they also create a direct financial target for increasingly capable AI systems.
Disclaimer
The information provided by Altcoin Buzz is not financial advice. It is intended solely for educational, entertainment, and informational purposes. Any opinions or strategies shared are those of the writer/reviewers, and their risk tolerance may differ from yours. We are not liable for any losses you may incur from investments based on the information provided. Bitcoin and other cryptocurrencies are high-risk assets; therefore, conduct thorough due diligence. Copyright Altcoin Buzz Pte Ltd.
The post Anthropic: AI Agents Find $4.6M Blockchain Exploits appeared first on Altcoin Buzz.
Â
This articles is written by : Nermeen Nabil Khear Abdelmalak
All rights reserved to : USAGOLDMIES . www.usagoldmines.com
You can Enjoy surfing our website categories and read more content in many fields you may like .
Why USAGoldMines ?
USAGoldMines is a comprehensive website offering the latest in financial, crypto, and technical news. With specialized sections for each category, it provides readers with up-to-date market insights, investment trends, and technological advancements, making it a valuable resource for investors and enthusiasts in the fast-paced financial world.