PureFi Announcements

🎓 PureFi at the 15th International Counter Fraud, Cybercrime and Forensic Accounting Conference Our CTO Mike Tiutin presented his research at the University of Portsmouth, UK, alongside 70+ experts from academia, law enforcement, government and industry. The conference covered AI-enabled crime, blockchain risks, cross-border fraud and forensic accounting, with speakers from the University of Cambridge, the UK Home Office and the Global Anti-Scam Alliance. 🧠 Can an AI agent do a better job than fixed rules at identifying suspicious crypto accounts? Getting a label wrong has consequences. A false positive can flag a legitimate customer. A false negative can let a criminal through. And that label feeds into alerts, risk scores and reports. Mike compared rule-based systems with an AI agent on two real-world tasks. Human experts reviewed the results, with 20 experts on the first task and 5 on the second. 1️⃣ Detecting permission-change attacks on TRON (UAP) These transactions change the keys controlling an account. The same action can be legitimate or malicious — the context makes the difference. • Rules: 15% false positives and ~60% of real attacks missed • AI agent: ~2% false positives and 99 out of 100 attacks detected The key detail: the agent handled only the ~80% of cases it was confident about, sending the hardest ~20% to a human reviewer. 2️⃣ Identifying crypto exchange wallets We expected this task to be easier. It wasn't. The rules produced few false positives but missed 62% of actual exchange wallets. The AI agent performed better. 📏 Confidence matters The rules had no way to express uncertainty — despite being right only about half the time. The AI agent's confidence scores aligned with expert judgments: when it was confident, experts agreed; when it said "50/50", experts were split too. ⏱ The trade-off Rules return results instantly. The AI agent takes around 2 minutes per account on the first task and 10 minutes on the second. ✅ The takeaway: use the right tool for each case. • Rules handle the easy cases fast. • The AI agent tackles the ambiguous ones. • Humans review the cases where the agent isn't confident. It's not about AI replacing everything. It's about building the right workflow for the task. *Limitations: two tasks, one blockchain, and expert agreement as the benchmark — not absolute ground truth.* Thanks to the Centre for Cybercrime and Economic Crime and the University of Portsmouth for hosting the conference. #PureFi #AML #Compliance #Cybercrime #CounterFraud #AI #Blockchain