AI Agents Reproduce $550M in Simulated Smart Contract Exploits, Anthropic Study Finds
Summary: Anthropic’s SCONE-bench shows AI models can autonomously exploit smart contracts, reproducing $550.1 million in simulated losses and discovering new zero-day vulnerabilities, while industry data highlights rising hack volumes in 2026.
Artificial intelligence agents have demonstrated the technical capacity to autonomously identify and exploit vulnerabilities in blockchain smart contracts, according to new benchmark results released by Anthropic. The findings indicate that frontier AI models can reproduce historic exploits and discover previously unknown security flaws, raising concerns about the future of decentralized finance (DeFi) security as automated hacking tools become more sophisticated.
SCONE-bench Findings
In December 2025, Anthropic published results from SCONE-bench, a comprehensive benchmark comprising 405 smart contracts that were exploited in real-world incidents between 2020 and 2025. The study evaluated the ability of AI agents to identify vulnerabilities and generate exploit scripts in sandboxed environments. The agents successfully reproduced exploits for 207 of the 405 contracts, resulting in a total of $550.1 million in simulated stolen funds. This metric provides a concrete lower bound for the economic harm these capabilities could enable if deployed maliciously.
To control for data contamination and ensure the agents were not simply memorizing known exploits, researchers tested the models on vulnerabilities that emerged after their training cutoffs. Claude Opus 4.5, Claude Sonnet 4.5, and GPT-5 were evaluated on these post-cutoff cases and collectively generated $4.6 million in simulated stolen funds. Anthropic emphasized that no actual funds were lost during these experiments, which were conducted entirely within local blockchain forks and simulators.
Beyond retrospective analysis, the agents were tested against 2,849 recently deployed smart contracts with no known issues. The analysis revealed two previously unknown zero-day vulnerabilities that could have resulted in $3,694 in simulated stolen funds. In one instance using GPT-5, the API cost for the experiment was $3,476, leaving a slim simulated profit. This demonstrates that profitable, autonomous exploitation is technically feasible even with current cost structures.
Rising Industry Losses
The advancement of AI-driven hacking tools coincides with significant financial losses in the crypto sector. Hacken, a blockchain security firm, reported that $1.18 billion was lost in 107 crypto hacks during the third quarter of 2026. Of this total, 48% was attributed to the exploitation of smart contracts. While Hacken did not specify the cause of these hacks, the volume of losses underscores the urgency for protocols to defend against increasingly automated threats.
Chainalysis estimates that approximately $36.7 million was stolen in 2026 from DeFi protocols that had not undergone smart contract verification. The firm notes AI-assisted reverse engineering as an emerging risk, though it does not attribute any specific thefts to AI agents at this time.
Benchmark Performance and Defense
Other industry benchmarks have further illustrated the evolving capabilities of AI agents. In February 2026, OpenAI and Paradigm released EVMbench, which showed that GPT-5.3-Codex achieved a 71% success rate in exploitation within a sandboxed environment. Later, in October 2026, CyberChainBench analyzed 541 exploit events across nine Ethereum-compatible blockchains. The best-performing configuration in that study achieved success rates of 37.5% for detection, 43.7% for exploitation, and 23.4% for patching.
An experiment by a16z on past DeFi price manipulation exploits found that success rates dropped from 50% to 10% when transactional data revealing the original exploit was restricted. However, when the AI was provided with high-level domain information, success rates jumped to 70%, suggesting that context significantly enhances AI exploitation capabilities.
Security experts advise that teams implement rigorous testing and monitoring to mitigate these risks. AI can be used defensively to test contracts prior to deployment, generate regression tests, and identify unusual on-chain behavior. However, human verification remains essential for signing off on changes. Best practices include continuing third-party audits, limiting access to sensitive information, and ensuring AI systems do not have permissions to transact without human oversight.