Top AIs Unready for High-Stakes Web3 Tasks

The newly established DMind Benchmark, rigorously peer-reviewed and accepted at KDD 2026, has evaluated 31 leading AI models, including GPT-5, Claude, and Gemini, assessing their capabilities across 3,543 expert questions in Web3 domains. The findings indicate a significant shortfall, with no tested AI system deemed ready for the most critical and financially consequential tasks within Web3.

Performance failures were most pronounced in safety-critical areas such as security vulnerability detection and token economics reasoning. These are precisely the domains where AI errors could result in catastrophic and irreversible financial losses, underscoring the current inadequacy of AI for unsupervised deployment in Web3's high-stakes environment.

The benchmark’s design emphasizes genuine multi-step reasoning, with adversarial fine-tuning experiments showing minimal gains, confirming that memorization cannot substitute for true understanding. Researchers highlighted that even top-performing models exhibit capability gaps unacceptable for real-world Web3 auditing or governance.

DMind Benchmark was developed by researchers from DMind AI, Zhejiang University, and Nanyang Technological University, with its paper accepted for presentation at KDD 2026 in Jeju, Korea, from August 9-13, 2026. The benchmark's open-source release on Hugging Face in April 2025 garnered significant traction, reaching the #1 trending position and accumulating over 9,650 downloads by January 2026.