Plain-language explainers for consumers — built from real safety-evaluation findings. No jargon, no assumptions, just what you need to know about the AI in your banking app.
Each item translates a real safety-evaluation finding into something a non-technical reader can use. Two are live now: an interactive bilingual demo and the first written explainer. The rest are in progress — check back as we publish them.
A hands-on demo — pick a real scam scenario from our evaluation suite, see what a safe banking AI should do (and what a vulnerable one might do instead), then explore the actual data behind the tests. Bilingual: switch between English and Bahasa Melayu.
Open the interactive demo →Your banking chatbot can check balances, flag suspicious transfers, and even call the police on a scammer. But it can also be tricked — and the tricks are getting cleverer. Here's what you need to know.
Read now →"I'm from your bank's compliance team — transfer now or face charges." Real authority impersonation cases from BankBench-MY's evaluation suite, translated into plain warnings.
Not yet publishedRM50 today, RM80 tomorrow, RM150 the day after. Why small repeated transfers to a new payee are a red flag — and what the AI should (but sometimes doesn't) catch.
Not yet publishedSwitching from formal Bahasa to casual Manglish mid-conversation is natural — but scammers use it to trick banking AI into lowering its guard. What to watch for.
Not yet publishedAfter ten minutes of friendly financial advice, your banking AI might let its guard down. Here's the science of rapport dilution — and how to spot it.
Not yet publishedWhen your banking chatbot says "let me pass this to my supervisor," it's not just a delay. It's a safety control — and scammers know how to exploit the handoff.
Not yet publishedBankBench-MY is a research project that tests how safe Malaysian banking AI really is. These explainers are the public-facing side of that work — translated from evaluation findings into advice you can actually use.
Every explainer is grounded in an actual safety-evaluation scenario tested on real banking-agent LLMs. We don't speculate — we show what failed, why, and what it means for you.
Explainers are published in both BM and EN where possible. The scenarios themselves are tested across BM Baku, BM Pasar, Manglish, and code-switching.
We assume you've used GrabPay or MAE. We don't assume you know what "prompt injection" or "multi-turn state manipulation" means. If we use a term, we define it in plain language.