Coinbase detailed a new LLM-based fraud detection agent for its Onramp product in an October 6, 2026 blog post, the first of a three-part “Owning Intelligence” series In a live online A/B tes
Coinbase published a blog post dated October 6, 2026, written by Yao Ma, Bo Lei and Xuwei Tan, describing how the company built, evaluated and post-trained a large language model agent to flag fraudulent transactions on its Onramp product, which lets users convert fiat currency into crypto.
According to the post, the agent achieved 9.6% higher performance, measured by F1 score, and 55% lower end-to-end latency compared with frontier general-purpose models the team benchmarked against during development. Rather than replacing Coinbase’s existing fraud infrastructure, the LLM agent runs downstream of established checks, providing what the company described as “an additional contextual review for selected transactions” flagged by earlier stages of the fraud pipeline.
Coinbase said it ran a live online A/B test comparing the new agent against its existing machine learning models and rule-based systems. The treatment group, which had the LLM agent active, saw 30% fewer fraudulent transactions and 22% less total fraud value compared with the control group relying solely on the prior detection stack, a result the company called a meaningful improvement over its previous fraud-prevention baseline.
The company also disclosed that projected savings from the improved fraud prevention were roughly three times the estimated cost of running the agent’s inference, framing the deployment as not just more effective but also economically favorable relative to the computing resources required to operate it. Architecturally, the system was built on Ray Serve, an open-source framework for scaling machine learning workloads, reflecting Coinbase’s broader infrastructure choices for deploying AI models in production.
The post is the first installment in Coinbase’s “Owning Intelligence” series, suggesting the exchange plans to detail additional internally built AI systems in future posts covering other parts of its product stack. The disclosure comes as crypto exchanges increasingly lean on AI-driven tools to combat fraud tied to onramp and offramp services, which remain common targets for bad actors looking to move illicit funds into or out of the crypto ecosystem through seemingly legitimate fiat conversions, making measurable fraud-rate reductions like the one Coinbase reported a closely watched benchmark across the industry.
This post first appeared in Coinbase’s New AI Fraud Agent Cuts Fraudulent Onramp Transactions by 30% in Live Testing
