GPT-5 and Claude AI Find Million‑Dollar Zero‑Day Flaws for Ethereum Smart Contracts
Frontier AI models — OpenAI GPT-5 and Anthropic Claude (Sonnet 4.5 / Opus 4.5) — don show say dem fit find and craft working exploits for Ethereum-compatible smart contracts by their own inside simulated environments. Joint project wey Anthropic red team and Machine Learning Alignment & Theory Scholars (MATS) run create SCONE-bench (Smart CONtracts Exploitation benchmark) and dem tested 405 contracts wey people don exploit before (2020–2025); across 10 models dem produce exploits for 207 contracts, simulating $550.1 million wey compromise. For contracts wey publish after model training cutoffs, top systems (Claude Opus 4.5, Claude Sonnet 4.5 and GPT-5) compromise 19 of 34 contracts, simulating about $4.6 million theft.
Extended tests target 2,849 recently deployed contracts wey no get reported bugs. Sonnet 4.5 and GPT-5 find two zero-day vulnerabilities wey nobody know before, and simulated profit near $3,694 (GPT-5 API cost for that test na $3,476). Claude architecture still show major efficiency gains, reduce token cost per successful exploit by ~70% compared to six months before and allow about 3.4× more attacks for the same compute budget. All experiments run for isolated simulated blockchains to prevent real-world harm.
Implications for traders: AI don lower cost and scale automated smart-contract exploits well, which increase systemic cyber risk to DeFi and on-chain assets — especially for contracts wey dem just deploy or never audit. Immediate trader actions include favour audited and battle-tested protocols, monitor exploit and on-chain flow alerts, tighten position sizing for exposure to newer projects, and watch for quick adoption of AI-powered security tools wey fit change detection and remediation timelines. Key SEO keywords: AI security, smart contracts, Ethereum, zero-day vulnerabilities, DeFi risk.
Bearish
Dis news na bearish for di cryptocurrencies an DeFi tokens dem wey dey affect because e dey show say systemic cyber risk don much higher. AI models wey dey reduce di cost an increase di scale of automated exploits dey raise di chance say attackers go fit successfully hit newer or unaudited smart contracts. Short‑term effects: more selling pressure an volatility for tokens wey tie to vulnerable or small‑cap protocols as traders dey de‑risk, increased on‑chain outflows to safer assets, an rapid price moves when exploit alerts show. Medium‑term effects: continued downward pressure on tokens of protocols wey no adopt robust audits or AI‑based defenses, higher risk premiums for newly launched projects, an possible re‑rating of security as a differentiator — wey go benefit well‑audited, blue‑chip DeFi platforms. Offsetting factors: faster adoption of AI‑powered defensive tools an pre‑deployment stress testing fit reduce exploit frequency over time, an white‑hat recoveries or insurance payouts fit limit final losses. Overall, di immediate net impact on price na negative for vulnerable protocol tokens but fit be neutral or positive for established, well‑secured tokens as capital shift toward safer venues.