GPT-5 and Claude AI Find Million‑Dollar Zero‑Day Flaws in Ethereum Smart Contracts
Frontier AI models — OpenAI’s GPT-5 and Anthropic’s Claude (Sonnet 4.5 / Opus 4.5) — have demonstrated the ability to autonomously find and craft working exploits against Ethereum-compatible smart contracts in simulated environments. A joint project by Anthropic’s red team and the Machine Learning Alignment & Theory Scholars (MATS) program created SCONE-bench (Smart CONtracts Exploitation benchmark) and tested 405 historically exploited contracts (2020–2025); across 10 models they produced exploits for 207 contracts, simulating $550.1 million in compromised value. For contracts published after model training cutoffs, top systems (Claude Opus 4.5, Claude Sonnet 4.5 and GPT-5) compromised 19 of 34 contracts, simulating about $4.6 million in theft.
Extended tests targeted 2,849 recently deployed contracts with no reported bugs. Sonnet 4.5 and GPT-5 found two previously unknown zero-day vulnerabilities, yielding simulated profits near $3,694 (GPT-5 API cost on that test was $3,476). The Claude architecture also showed major efficiency gains, cutting token cost per successful exploit by ~70% versus six months earlier and enabling roughly 3.4× more attacks for the same compute budget. All experiments ran in isolated simulated blockchains to prevent real-world harm.
Implications for traders: AI materially lowers the cost and scales automated smart-contract exploits, increasing systemic cyber risk to DeFi and on-chain assets — especially for recently deployed or unaudited contracts. Immediate trader actions include favouring audited and battle-tested protocols, monitoring exploit and on-chain flow alerts, tightening position sizing for exposure to newer projects, and watching for rapid adoption of AI-powered security tools that could change detection and remediation timelines. Key SEO keywords: AI security, smart contracts, Ethereum, zero-day vulnerabilities, DeFi risk.
Bearish
This news is bearish for the affected cryptocurrencies and DeFi tokens because it highlights materially higher systemic cyber risk. AI models lowering the cost and increasing the scale of automated exploits raises the probability of successful attacks against newer or unaudited smart contracts. Short-term effects: heightened selling pressure and volatility for tokens tied to vulnerable or small-cap protocols as traders de-risk, increases in on-chain outflows to safer assets, and rapid price moves when exploit alerts surface. Medium-term effects: continued downward pressure on tokens of protocols that fail to adopt robust audits or AI-based defenses, greater risk premiums for newly launched projects, and potential re-rating of security as a differentiator — benefiting well-audited, blue‑chip DeFi platforms. Offsetting factors: accelerated adoption of AI-powered defensive tools and pre-deployment stress testing could reduce exploit frequency over time, and white-hat recoveries or insurance payouts may limit ultimate losses. Overall, the immediate net impact on price is negative for vulnerable protocol tokens but may be neutral or positive for established, well-secured tokens as capital shifts toward safer venues.