Market desk Bitcoin Ethereum Altcoins DeFi Stablecoins Markets & Trading

AI Agents Are Finding and Exploiting Smart Contract Vulnerabilities at Increasing Rates

Recent research shows AI systems can identify and exploit smart contract vulnerabilities in controlled environments, with success rates between 10% and 71% depending on available data. Security teams are urged to adopt rigorous testing and monitoring practices.
28 minutes ago 1 views
AI Agents Are Finding and Exploiting Smart Contract Vulnerabilities at Increasing Rates

How AI Agents Identify Vulnerabilities

AI agents are becoming more effective at finding smart contract weaknesses by iterating through attack attempts, analyzing dependencies, and using local blockchain forks and decompilation. This approach can expose vulnerabilities that pattern-based scanners typically miss, including access control bugs, accounting errors, oracle-related issues, and unexpected contract interactions.

In December 2025, Anthropic released findings from SCONE-bench, a benchmark of 405 smart contracts that were exploited between 2020 and 2025. AI agents successfully reproduced exploits for 207 contracts, representing $550.1 million in simulated stolen funds. When tested on previously unknown vulnerabilities published after the models' training cutoff, Claude Opus 4.5, Sonnet 4.5, and GPT-5 generated $4.6 million in simulated losses. In tests on 2,849 recent contracts with no known issues, researchers discovered two previously unknown vulnerabilities that could have resulted in $3,694 in simulated losses.

Real-World Hacking Trends

Hacken reported that $1.18 billion was lost across 107 crypto hacks in the third quarter of 2026, with 48% stemming from smart contract exploits. Chainalysis estimated that approximately $36.7 million was stolen in 2026 from DeFi protocols that had not undergone smart contract verification, noting AI-assisted reverse engineering as an emerging risk.

In an a16z experiment on 20 past DeFi price manipulation exploits, success rates fell from 50% to 10% when researchers restricted AI access to transactional data revealing the original exploit. When given high-level information about the DeFi domain, success rates rose to 70%.

Benchmark Results

In February 2026, OpenAI and Paradigm released EVMbench, a benchmark for evaluating AI agents' vulnerability detection, patching, and exploitation capabilities. GPT-5.3-Codex achieved a 71% success rate in exploitation within a sandboxed environment and 33.3% for GPT-5.

CyberChainBench, released in October 2026, analyzed 541 exploit events across 9 Ethereum-compatible blockchains. The best-performing configuration achieved 37.5% success for detection, 43.7% for exploitation, and 23.4% for patching. In simulation, the top-performing agent generated $57.4 million in simulated profits from exploits on 200 events occurring after the training cutoff.

Recommended Security Practices

Development teams should implement rigorous testing and monitoring to stay ahead of potential AI-driven threats. Recommended measures include using AI to test contracts before deployment, generating regression tests, and identifying unusual on-chain behavior, with humans verifying all potential issues before approval.

Additional security practices include third-party audits, limiting access to sensitive information, setting up alerts and response plans, restricting wallet balances, and ensuring AI systems require human oversight before executing transactions. Smaller protocols may face increased exploit attempts as AI agents can test attacks against many protocols efficiently.

Market snapshot

Top cryptocurrency prices

Explore all prices
BitcoinBTC $63,578.23-1.68% EthereumETH $1,805.75-0.32% Tether USDUSDT $0.9996-0.02% BNBBNB $589.02-0.54% USDCUSDC $1.00-0.01% XRPXRP $0.9819-0.33% SolanaSOL $74.99+0.08% TRONTRX $0.3126+1.35% HyperliquidHYPE $68.26-0.95% ZcashZEC $943.67-1.59%
Prices by Coinranking. Informational only.