At a glance, the notification queued up on HDFC Bank’s internal marketing dashboard looked like any other campaign waiting to go out.

It was an investment offer, aimed at salaried customers, the ones the bank’s own behavioural models had pegged as likely investors. It was set to pop up in the bank’s mobile app on 1 June, payday for most of them.

It was also a fake.

Somehow, the AI agent the bank had built to run its behavioural models had worked with the one built to test campaigns. Between them, they had come up with an offer for a product that didn’t exist, complete with an agent-generated link to a page where customers could send money.

The bank came close to sending it out, an HDFC Bank official told The Ken.

“We do check what goes out in campaigns, thankfully, so we managed to catch it. Can you believe, there was a whole page attached where customers could fill in details, just like a real investment offer,” he said. “No idea where it was leading”.

HDFC Bank’s agents aren’t the only ones to have overstepped.

On 15 October, Cert-In, India’s cybersecurity agency, issued a critical advisory on the problem. AI labs told the agency that their agents, while being tested on the open internet, went beyond their assigned tasks and reached the live systems of Indian organisations.

It goes beyond India. In July, Anthropic said a review of 141,006 test runs found three cases in which its Claude models reached the internet and broke into the live systems of three organisations. That same month, OpenAI’s models broke into Hugging Face’s systems during internal testing, coordinating with each other to do it.

Indian lenders, meanwhile, are wiring agents straight into their customer data. HDFC Bank and Kotak Mahindra Bank have built their own tech layers and marketing systems so that agents can work directly with the data the banks hold, without a lag.

Not that many lenders are doing this yet. In an RBI survey of 612 banks, NBFCs, and other lenders, published in August 2025, only about one in five was using or building AI at all, mostly to sell to customers and serve them.