A landmark research project from Anthropic has demonstrated that frontier AI models can autonomously discover and exploit vulnerabilities in real-world smart contracts, collectively identifying exploits worth $4.6 million. The findings, published on December 2, 2025, reveal both the transformative potential and the significant risks at the intersection of artificial intelligence and blockchain technology.
The Synergy
The research, conducted through the MATS and Anthropic Fellows program, evaluated leading AI models including Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 against a newly constructed benchmark called SCONE-bench (Smart CONtracts Exploitation benchmark). This benchmark comprises 405 smart contracts that were actually exploited between 2020 and 2025, providing a rigorous testing ground for evaluating AI capabilities in smart contract analysis.
The results were striking. When tested against contracts exploited after the models’ knowledge cutoff dates — ensuring they could not simply recall known vulnerabilities — the AI agents developed functional exploits with a combined value of $4.6 million. This establishes a concrete lower bound for the economic impact that AI-powered vulnerability discovery can achieve.
AI Use Cases in Web3
The implications for the Web3 ecosystem are profound and dual-natured. On the defensive side, AI agents could revolutionize smart contract auditing. Current manual auditing processes are expensive, time-consuming, and often miss subtle vulnerabilities. An AI-powered auditing tool could analyze every deployed contract on a blockchain in real-time, identifying vulnerabilities before attackers can exploit them.
Several projects are already building in this direction. The research demonstrated that AI agents can identify not just known vulnerability patterns but also novel attack vectors. In a forward-looking experiment, the researchers tested both Sonnet 4.5 and GPT-5 against 2,849 recently deployed contracts with no known vulnerabilities. Both agents uncovered two novel zero-day vulnerabilities, producing exploits worth $3,694.
Notably, GPT-5 achieved this at an API cost of $3,476 — meaning the operation was nearly profitable even in this exploratory mode. This demonstrates as a proof-of-concept that profitable, real-world autonomous exploitation is technically feasible, a finding that dramatically raises the stakes for blockchain security.
Data Privacy Implications
The intersection of AI and blockchain security raises important questions about data privacy and responsible disclosure. The Anthropic researchers emphasized that all exploit testing was conducted exclusively in blockchain simulators — never on live networks. This responsible approach sets an important precedent for how AI security research should be conducted in the blockchain space.
However, the same capabilities could be weaponized by malicious actors. The research found that exploit revenue from simulated AI-generated attacks has roughly doubled every 1.3 months over the past year. This exponential growth in AI cyber capabilities means that the window between vulnerability introduction and exploitation is shrinking rapidly.
The Innovation Frontier
The convergence of AI and crypto is also creating entirely new categories of applications. Crypto AI agents are evolving from simple automated scripts into intelligent, adaptive entities that can navigate the complexities of the Web3 ecosystem. Projects like Fetch.ai are building universal agent frameworks for decentralized digital economies, while SingularityDAO is deploying AI-managed hedge funds on-chain.
The Anthropic research suggests that the next frontier is AI-versus-AI security: AI agents defending smart contracts against AI-powered attacks. This arms race will drive demand for more sophisticated on-chain monitoring, real-time vulnerability detection, and automated patch deployment — all areas where blockchain-native AI systems have natural advantages.
Concluding Thoughts
The $4.6 million figure from Anthropic’s research is not just a number — it is a preview of how AI will fundamentally reshape blockchain security. The technology that enables an attacker to autonomously discover a vulnerability worth millions is the same technology that can protect protocols before those vulnerabilities are ever exploited. The question is not whether AI will transform blockchain security, but whether the defensive applications will scale fast enough to outpace the offensive ones. For the crypto industry, investing in AI-powered security is no longer optional — it is an existential imperative.
Disclaimer: This article is for informational purposes only and does not constitute financial or investment advice. Always conduct your own research before making any financial decisions.
Bridge security is still the weakest link in the ecosystem
The cost of a security breach always exceeds the cost of prevention
Bug bounties are the most cost-effective security investment
The amount of DeFi exploits is still way too high
Real-time monitoring tools are getting better at catching exploits early
4.6M is the academic lower bound. a dedicated team with compute and time would find 10x that. the dual use risk is genuinely scary
$4.6M in novel exploits from AI agents. and thats a lower bound. imagine what a well funded team could find that these academic researchers didnt
SCONE-bench with 405 real exploited contracts is a legit benchmark. $4.6M in novel exploits means AI auditing tools could save protocols millions if deployed defensively
SCONE-bench is actually rigorous. 405 real exploits tested against models post knowledge cutoff. this is how you properly evaluate AI security capabilities
bug_finder_ testing against post-cutoff contracts was the right call. proves the models actually reason about vulnerabilities instead of memorizing CVEs
the dual use problem is the real issue. same AI that finds bugs for auditors can find bugs for attackers. its an arms race and the attackers are faster
the dual use problem is the scariest part. same model that finds a bug for an auditor can write the exploit for an attacker. whoever moves fastest wins