Coinbase said its purpose-built AI Agent outperformed a leading commercial AI system in fraud detection while responding far faster to payment reviews, cutting fraudulent transactions by 30% and fraudulent transaction value by 22% during testing on the company's Onramp service.

The agent is built on Qwen3.5-9B, an open-weight model that Coinbase engineers fine-tuned on proprietary internal transaction data. Chief Executive Brian Armstrong discussed the results on October 12, following a three-part engineering series the company published the previous week.

"We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase , and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case). I think we'll see more of this pattern inside… pic.twitter.com/x1gort3N6P"

— Brian Armstrong (@brian_armstrong) October 11, 2026

In the post, Armstrong suggested the outcome pointed to a broader pattern in which open-weight models fine-tuned on proprietary data can beat frontier systems on narrow, specialized tasks at lower cost and higher speed.

AI Reviewer Reduces Onramp Fraud

Coinbase initially added the AI reviewer alongside its existing fraud rules and machine-learning checks on Onramp, the company's service for purchasing cryptocurrency with fiat funds. In an online experiment, the company recorded 30% fewer fraudulent transactions and a 22% reduction in fraudulent transaction value.

Rather than judging each payment in isolation, the agent reviewed sequences of recent payments to assess suspicious activity. That allowed it to stop a payment that had passed earlier checks, although it lacked permission to reverse a block that was already in place. The model returned risk classifications only, with application code applying the final transaction decisions — an arrangement the company said strengthens transaction security while maintaining clear limits on automated fraud prevention decisions.

Internal Testing Supports Smaller Fine-Tuned Model

Coinbase compared several models using a test set of 16,140 transactions involving 7,293 users, including 813 confirmed fraud cases. The company found that some newer general-purpose models missed more fraud despite improvements in the accuracy of the transactions they flagged.

Engineers then adapted Qwen3.5-9B, a nine-billion-parameter open-weight model, using Coinbase's historical fraud records. According to the company's report, training ran on a single graphics processor and cost less than $100.

Against Anthropic's Opus 4.5, the adapted model improved its F1 score by 9.6 percentage points. Coinbase also reported a 55% lower median response time, with its model taking 0.683 seconds per request compared with 1.515 seconds for the commercial system, based on production measurements.

The findings come from company-run tests rather than an independent assessment of how the models perform across other payment platforms. As a result, the reported comparison is specific to Coinbase's transaction-review workflow and test conditions rather than a general ranking of the models.

Clear Limits on Automated Decisions

Coinbase said its model requires continued testing as fraud patterns change, with established application rules controlling how the system acts. The approach aligns with a warning from security firm CertiK last week that autonomous agents need clear limits and human accountability. For the system described by Coinbase, monitoring whether classifications remain effective as fraud patterns change will therefore remain part of operating the tool alongside those application rules.

The use of specialized AI for fraud prevention comes as blockchain developers explore other security applications. An AI-assisted discovery of an XRP Ledger flaw recently led to a fix for a vulnerability that could have allowed unauthorized token creation. Separately, Vitalik Buterin's warning about AI and wallet security raised concerns that advanced systems could also threaten existing cryptographic protections, while a phishing incident involving Coldcard's social account showed that crypto security faces threats beyond transaction fraud.