Coinbase Slays Fraud Detection Using A Compact Model 🖤
In their October 12 post Coinbase CEO Brian Armstrong spilled that they fine-tuned Qwen3.5-9B on internal data and it crushed frontier models on fraud prevention at way lower cost and speed. The real tea is that specialist models win on narrow tasks whether solo or teaming up. 💀 The exchange broke it down in their Owning Intelligence series showing how they added an AI risk agent to their machine learning setup for Onramp. This lets users snag crypto via partners even in guest mode. An experiment cut fraudulent transactions by 30 percent and fraud value by 22 percent. The agent scanned recent transaction sequences spotting sneaky patterns a single payment check would miss. It can block shady moves that passed earlier filters but never reverse a prior block. They benchmarked fresh closed-source models on exactly 16,140 transactions from 7,293 users including 813 confirmed fraud cases. Newer Opus Sonnet and GPT versions actually flopped harder on multiple fraud metrics than older drops. GPT-5.6 boosted precision yet tanked recall so it missed more fraud overall. That pushed them toward open-weight models they could tweak directly instead of API-only vibes. 🔮AI Autonomy Sparks Some Sinister Warnings 🚀
Coinbase stays cautious with clear limits on their agent since it only returns a risk score and app code enforces the final call. It blocks allowed transactions but cannot override existing blocks. Certik flagged risks around autonomous security agents potentially causing bad outputs and less human oversight so named owners and hard caps remain key lol. 😈 This whole approach feels kinda iconic for DeFi safety right now. 💅Just another echo from the void by iconofsin.eth 💖
Maybe you like what i'm doing here and wanna support me via the ethereum blockchain: iconofsin.eth 💖
