Coinbase says a specialized AI agent cut fraud in a live test of its Onramp product. The company fine-tuned an open-weight model on its own transaction data and found it beat a leading closed-source model on fraud detection, with better scores and lower serving latency.
Specialized model beats general-purpose rivals
The exchange laid out the work in a three-part technical series called Owning Intelligence. CEO Brian Armstrong said in an October 12 post that Coinbase fine-tuned Qwen3.5-9B using internal data. He wrote that the model “outperformed frontier models, at much lower cost and speed” for fraud prevention.
Armstrong’s point was not that small models always win. He argued that performance depends on the task. General-purpose models are strong across many jobs, he said, but specialist models can do better on narrow work.
Coinbase first added an AI risk agent to its existing machine-learning model and rules engine for Onramp. The product lets users buy crypto through partner apps, including guest checkout without a Coinbase account.
In an online experiment, the group with the extra AI review saw 30% fewer fraudulent transactions and 22% less fraud value. The agent looked at recent transaction sequences to spot patterns that a single payment might not show. It could block a transaction that passed earlier checks, but it could not undo an upstream block.
Benchmark shows mixed results for closed models
Coinbase then tested newer closed-source models against a fixed benchmark of 16,140 transactions from 7,293 users. That set included 813 confirmed fraudulent transactions. The company reported that newer Opus, Sonnet, and GPT versions performed worse on several fraud-detection measures than their predecessors. GPT-5.6 improved precision but had lower recall, so it caught a smaller share of fraudulent transactions.
That result pushed Coinbase toward an open-weight model. With open weights, the company could adapt parameters directly instead of reaching the model only through a provider’s API. That matters for fraud, where each firm’s data and risks differ.
Autonomy concerns remain
The move comes as security firms debate how much freedom AI agents should have in financial crime work. Certik warned last week that more autonomous security agents could bring risks from incorrect outputs and less human scrutiny. It recommended clear limits on autonomous decisions and a named owner for every agent.
Coinbase’s agent appears to work inside those limits. It returns a risk score, and application code applies the decision policy. It can block a transaction that earlier checks allowed, but it cannot override a block that has already been made.
That setup may be the real lesson. Coinbase is not claiming AI can replace human oversight. It is testing where a focused model helps and what controls need to stay in place. The 30% drop in fraud is notable, but it comes from one experiment, not a guarantee for every crypto platform.
