Coinbase (COIN) AI Agent Cuts Fraudulent Payments 30% on Bitcoin (BTC) Exchange

Qwen3.5-9B Deployed on Onramp

Coinbase has cut fraudulent transactions by 30% after putting an AI risk agent into live payment checks, a result the company reports from an experiment on Onramp, its crypto purchase payment solution. The model behind the deployment is Qwen3.5-9B, an open-weights model that Coinbase fine-tuned on its own transaction data instead of renting a closed model through an external provider. In the online experiment, the agent ran as an added verification stage for one user group, and in that group anomalous transaction count fell 30% while measured fraud losses fell 22%. CEO Brian Armstrong said the specialized model completed anomaly-detection work successfully at a much lower cost and faster processing speed than state-of-the-art frontier models. His framing: a general-purpose model handles broad tasks well, but a small model purpose-trained for a single domain proves stronger inside that domain. The deployment matters beyond one company metric. Onramp sits on the purchase path through which retail buyers take on Bitcoin (BTC) price exposure, so screening quality on these rails shapes the Bitcoin security layer that ordinary users touch first. Unlike the legacy rules engine, the agent does not judge a payment in isolation; it weighs the recent sequence of transactions and flags abnormal patterns that fixed rules tend to miss. The rollout also lands amid a wider industry debate over AI in financial security, after the security firm CertiK warned about malfunction and absent human oversight in autonomous agents. Coinbase engineered its answer into the architecture itself: the agent scores risk and nothing more, application code executes the final policy, and the agent may block a payment the existing system approved but cannot release one blocked upstream.

16,140-Payment Benchmark

Underneath the headline number sits a benchmark Coinbase ran against the newest commercial closed models. The test set covered 16,140 payments built on data from 7,293 users and included 813 confirmed fraudulent transactions. Coinbase cross-validated the latest frontier models from the Opus, Sonnet and GPT families against that base, and several scored below older generations on certain fraud metrics, a reversal of the usual upgrade curve. GPT-5.6 illustrates the pattern: precision improved, yet recall, the share of real fraudulent transactions the model actually catches, declined. For a security team, a model that misses live fraud is unusable regardless of benchmark polish, and that gap is what pushed Coinbase toward open weights, where it can adjust every parameter directly instead of depending on an external provider's API roadmap. The market read on the exchange itself stays cautious. COINOTAG data shows Coinbase (COIN) changing hands at $182.49, down 0.37% over 24 hours, pinned just beneath the strongest resistance level in our composite at $182.19, which scores 100/100 against the EMA 20, EMA 50 and EMA 200 cluster. The nearest strong support sits at $175.89, also scored 100/100, while the daily trend still reads a downtrend with RSI at 47.7 and a bearish MACD signal. Derivatives positioning around the stock stays modest, with perpetual futures open interest at $24,698,582 and a funding rate of 0.0039% per interval.

Recall Beat the Frontier Giants

COINOTAG's reading: the benchmark, not the headline percentage, is the load-bearing result. With 16,140 payments, 7,293 users and 813 confirmed fraud cases, Coinbase compared models on recall, the one metric that decides whether fraud reaches a user, and the frontier brands lost there to a small open-weights model tuned in-house. That logic travels across the wider Bitcoin (BTC) market. Retail accounts that simply HODL through volatility gain the most from better screening at the purchase layer, institutional vehicles such as the spot ETF sit on the same exchange security stack for custody, and even a crypto whale routing eight-figure purchases benefits from sequence-level anomaly detection. The direction also rhymes with the AI revenue pipelines now forming around mining, where Castle Labs has priced Bitcoin miners' AI revenue share at 22%. As Armstrong framed it, the goal was a specialized model doing the anomaly-detection work of the frontier giants at lower cost and faster speed, and the 30% fraud reduction is the evidence he offered for it.