Coinbase AI Agent Reduced Fraud by 30% in Onramp Experiment

Coinbase reported that its AI agent reduced the number of fraudulent transactions by 30% in an experiment within the Onramp program. Coinbase CEO Brian Armstrong explained that the company further trained the Qwen3.5-9B model on internal data and found that it "outperformed state-of-the-art models at much lower costs and higher speeds" in the narrow task of fraud prevention.

Armstrong said, "State-of-the-art models are amazing generalists. But increasingly, you will see specialized models outperform them in more narrow tasks—both as part of a team and independently."

The experiment with the AI agent showed 30% fewer fraudulent transactions and a 22% lower volume of fraudulent operations in the group where additional AI verification was applied. The agent analyzed sequences of recent transactions to identify patterns not noticeable in a single payment. It could block a transaction that passed initial checks but could not reverse a previously made block.

Coinbase also tested new versions of closed models against a fixed benchmark of 16,140 transactions from 7,293 users, including 813 confirmed fraudulent ones. New versions of Opus, Sonnet, and GPT showed worse results across several fraud detection metrics than their predecessors. GPT-5.6 improved accuracy but showed lower recall, meaning it identified a smaller share of fraudulent transactions. This led to a shift towards an open model, whose parameters can be directly adapted rather than solely through a provider's API.

Coinbase's approach fits into a broader discussion about the role of AI agents in combating financial crimes and in the security sphere. Last week, Certik warned that the rise of autonomous security agents carries risks due to erroneous conclusions and reduced human oversight, recommending clear limits on autonomous solutions and an appointed responsible person for each agent. However, Coinbase's agent operates within set boundaries: it returns a risk assessment, and the decision is applied by the application code. It can block a transaction that previous checks missed but cannot reverse an already made block.