Newer AI models missed more payment fraud in Coinbase’s benchmark

Coinbase's benchmark finding that newer AI models missed more payment fraud than earlier versions, despite GPT’s precision gains, highlights a regression in fraud detection performance across three model upgrades. For website owners, especially those running e‑commerce or financial services sites, this signals that relying on the latest AI‑driven anti‑fraud tools may not automatically improve security. The benchmark shows that newer models can overlook suspicious transactions, increasing exposure to chargeback attacks and fraudulent bot traffic. Owners must therefore reassess any recent AI fraud filters they’ve implemented and verify that their detection rates have not slipped, rather than assuming newer equals better. Actionable steps: 1. Use llms.txt to explicitly list the AI models you trust for transaction monitoring, blocking unverified or experimental models that may under‑perform. 2. Deploy llms‑central’s bot‑traffic analytics to compare fraud detection rates before and after each AI model upgrade, flagging any rise in false negatives. 3. Set up automated alerts in your llms.txt configuration to trigger secondary verification (e.g., CAPTCHA or manual review) when the benchmark‑identified “missed fraud” pattern is detected in real‑time traffic.