Banks are fighting AI scams with AI tools that fail independent checks
A survey of Indonesian banking leaders puts social engineering at the top of the fraud list, a preprint finds financial agents failing verification in every session, and U.S. regulators are asking banks about audit trails.
- In a survey by BioCatch of 100 fraud, anti-money-laundering and compliance leaders at Indonesian banks, social engineering was named the top fraud concern, with 43% reporting increasing fraud attempts. Self-reported m:…1707
- Malaysian police arrested 35 foreign nationals in Penang on suspicion of running a romance and gold-investment scam centre targeting Chinese women via Telegram, seizing 119 mobile phones and three laptops. m:…9678
- In a preprint, the FINCOSMOS evaluation found that across 150 sessions, no session was accepted, and every session signed off by the simulated user failed a program check. Self-reported m:…1813
- Uptiq says its customers have seen a 36% reduction in financial spreading time and 41% faster underwriting, while U.S. regulators are questioning banks on AI agent limits, audit trails and human involvement. Self-reported m:…4627
The fraud bankers name first
Banks are adopting artificial intelligence faster than they can check it, and the evidence this week comes from both sides of the problem. A survey by BioCatch of 100 fraud-management, anti-money-laundering and compliance leaders at Indonesian banks identifies social engineering as their top fraud concern. Of those surveyed, 43% report increasing fraud attempts at their organisations, and 42% report increasing fraud losses. More than half, 55%, are actively evaluating, procuring or implementing new fraud-prevention solutions. m:…1707
The gap the bankers name is specific. Half of the surveyed leaders identify coached user journeys, cases where a victim is talked through a transaction by a scammer, as a major fraud-detection gap, and 19% rank it as their single greatest weakness. At the same time, more than four in five consider their existing fraud controls effective, and only 34% estimate their institution's annual fraud losses exceed $10 million. Confidence in the controls and a named hole in them sit side by side. m:…1707
The scam centres behind the numbers
The hole is not theoretical. Malaysian police arrested 35 foreign nationals in Bayan Lepas, Penang, on suspicion of operating an online love-scam and gold-investment scam centre that had been running since September 2026, targeting Chinese women using Telegram. Police seized 119 mobile phones and three laptops, with the seized equipment valued at approximately RM60,000. Tookitaki, which reported the bust, says romance scams and investment scams are increasingly reinforcing each other as part of a single scam journey, and that organised centres split recruitment, communication, collection and movement of proceeds between different people. m:…9678
Tookitaki's advice to banks follows from how the centres work: watch for a customer's behaviour changing under social engineering, and check whether accounts receiving payments from many unrelated people, with no clear economic reason, belong to a broader scam network. Banks should also monitor for sudden increases in transaction frequency and for beneficiaries linked to earlier scam complaints. m:…9678
The tools have a verification problem
The tools banks are buying to keep up have their own verification problem. In a preprint, researchers behind FINCOSMOS, an evaluation framework for financial AI agents, report that across 150 sessions, no session was accepted under the framework, and every session signed off by the simulated user failed a program check. Only 25 of 120 sessions delivered the investor's strategy, and human experts attributed 88.33% of failures to the agents. An agent with persistent memory learned all ten tasks after receiving feedback from the framework, which suggests the checking, not just the agent, is part of what is missing. m:…1813
The commercial side of the same story is moving anyway. Uptiq, which sells AI financial spreading for commercial lending, says its customers have seen a 36% reduction in spreading time, 41% faster underwriting, and up to three times as many deals per analyst, with 95% or better extraction accuracy and low-confidence fields routed to analysts. The company says U.S. regulators are questioning banks on AI agent limits, audit trails and human involvement, and notes that Federal Reserve guidance SR 26-2 and OCC Bulletin 2026-13 replace SR 11-7 with a risk-based approach to model risk management, while generative and agentic AI models are currently excluded from the revised guidance's scope. m:…4627
A control gap, not a technology gap
Put together, the week describes a control gap rather than a technology gap. The scams work by coaching a real person through a real payment, which is why coached journeys defeat controls built to spot fake accounts. The agents fail because they produce convincing reports that mask wrong data or flawed backtests; the preprint finds that financial agents working on the open web can fail without any visible error. And the regulators are asking about audit trails and human involvement while the newest guidance leaves generative models outside its scope. Banks are buying the tools. The open question is whether anyone can prove what the tools did. m:…1707 m:…1813 m:…4627
What is still unresolved
- Whether banks close the coached-journey gap, given that most surveyed leaders consider their current controls effective.
- Whether financial agents improve under program-based verification, or whether the FINCOSMOS failure rates hold for other agent designs.
- How U.S. regulators will treat generative and agentic AI models that sit outside the revised model risk guidance.
- Whether the Penang bust disrupts the wider romance and investment scam networks, or only one centre.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.