Coinbase AI Fraud Detection Cuts Onramp Fraud by 30%

Coinbase says its AI fraud detection system cut fraudulent transactions on its Onramp crypto-buying service by 30% and reduced the dollar value of fraud by 22% in an online A/B test. Onramp lets users buy crypto through partner apps, including via guest checkout. Coinbase first added an AI risk agent to its existing machine-learning checks and rules engine. The agent reviews transaction sequences and returns a risk rating for payments that passed earlier checks; application code makes the final decision. It cannot reverse a block made by an upstream system. The company then fine-tuned the open-weight Qwen3.5-9B model on proprietary transaction data. On a benchmark of 16,140 transactions from 7,293 users, including 813 confirmed fraud cases, Coinbase says the model outperformed Anthropic’s Opus 4.5 on precision, recall, F1 score and dollar-weighted recall. It reported a median response time of 0.683 seconds—55% faster than Opus 4.5—and training costs of less than $100 on a single GPU. Coinbase also said some newer frontier-model versions scored worse than earlier ones on parts of the benchmark; GPT-5.6 improved precision but had lower recall. The results suggest task-specific AI fraud detection could help crypto platforms reduce fraud losses, latency and operating costs. However, Coinbase’s proprietary benchmark has not been independently verified and covers one product and use case, so the findings do not establish broader performance across the industry.