TLDR
- Coinbase AI Agent reduced fraudulent transactions by 30% and fraud value by 22% during testing.
- Coinbase’s fine-tuned Qwen3.5-9B model outperformed leading closed-source AI models in fraud detection.
- The specialized AI model achieved 55% lower response latency than Anthropic’s Opus 4.5.
- Coinbase tested AI models against 16,140 transactions involving 7,293 users and 813 confirmed fraud cases.
- Coinbase AI Agent strengthens transaction security while maintaining clear limits on automated fraud prevention decisions.
Coinbase says its tailored fraud detection model has outperformed a commercial AI system, while responding faster to payment reviews. The Coinbase AI Agent uses internal transaction records to assess suspicious activity on its Onramp service.
CEO Brian Armstrong discussed the results on October 12, following a three-part engineering series issued last week. The findings come from company-run tests rather than an independent assessment of performance across other payment platforms.
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
Coinbase AI Agent Cuts Onramp Fraud
Coinbase initially added an AI reviewer to fraud rules and machine-learning checks. In an online experiment, the company recorded 30% fewer fraudulent transactions and a 22% reduction in fraudulent transaction value.
The agent reviewed sequences of recent payments instead of judging each transaction alone. It could stop a payment that passed earlier checks, but it lacked permission to reverse an existing block.
The use of specialized AI 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.
Internal Testing Supports Smaller Model
Coinbase compared several models using 16,140 transactions involving 7,293 users, including 813 confirmed fraud cases. Some newer general-purpose models missed more fraud despite improvements in the accuracy of flagged transactions.
Engineers then adapted Qwen3.5-9B, a nine-billion-parameter open-weight model, using Coinbase’s historical fraud records. Training ran on a single graphics processor and cost less than $100, the company’s report said.
Against Opus 4.5, the adapted model improved its F1 score by 9.6 percentage points. Separately, Vitalik Buterin’s warning about AI and wallet security raised concerns that advanced systems could also threaten existing cryptographic protections.
Faster Reviews With Clear Decision Limits
Coinbase reported a 55% lower median response time than Opus 4.5. Its model took 0.683 seconds per request, compared with 1.515 seconds for the commercial system, based on production measurements.
The Coinbase AI Agent returned risk classifications, while application code applied transaction decisions. Meanwhile, a phishing incident involving Coldcard’s social account showed that crypto security also faces threats beyond transaction fraud.
Security firm CertiK warned last week that autonomous agents need clear limits and human accountability. Coinbase said its model requires continued testing as fraud patterns change, with established application rules controlling how the system acts.
