Guide · AI security · as of 17 November 2026

Hackers redirected approved loans through a trusted address. Five checks.

On 16 November 2026 The Indian Express reported that twelve lenders in Delhi and its suburbs lost over Rs 65 crore in three months. Rs 65 crore is 650 million rupees. The money was loans that had been approved and were waiting to be paid out. Attackers changed the bank account and sent the payout order themselves. Police told the paper that in one case an AI attack tool sent the commands. The lenders’ systems accepted those orders because they came from a network address on a trusted list. This guide explains why that list is not a lock, and gives five checks for any business whose software can send money: loans, salaries, supplier payments or refunds.

What the police describe

Seven days inside, then the payout.

The account is the newspaper’s, from Delhi Police officers. The attackers put malware into the public side of a lender’s app, the screens where customers apply for loans. It stays for seven to eight days. In that time it reads which loans are approved but not yet paid, and learns the exact commands the system uses to pay one out. Then the borrower’s account number is replaced with an account the attackers control. The twelve cases are part of more than 150 complaints from lenders across India in the same period, with over Rs 300 crore taken. Police arrested 20 people, all linked to the bank accounts that received the money, and stopped Rs 18 crore before it left those accounts. A deputy commissioner in the police cyber unit said: “These incidents are largely attributable to weaknesses in the public-facing cyber systems” of the lenders.

Why the trusted list failed

A shared address is not a login.

An IP address is the number a computer sends from. Many payment systems keep an allow-list of addresses, often called a whitelist, and act only on requests from that list. The lenders’ lists held the addresses of the cloud data centres where their own websites run. An officer explained the gap: “most of these cloud services are publicly available, so anyone can open an account on them and use a specific set of IP addresses.” So the attackers rented servers from the same cloud and sent the payout request from there. In the officer’s words, it “bypasses the NBFC firewall because of the ‘whitelisted’ IP address and reaches the bank, getting instant approval.” In one case police counted more than 317 transfers sent from at least 165 addresses. An allow-list answers “where did this come from?” It does not answer “who sent it?” A cloud’s address range belongs to every customer of that cloud.

The same mistake inside one machine: your AI server trusts localhost

Where AI comes in

Same door, faster hands.

Police told the paper that in an October case, where a lender lost Rs 3 crore, “a Chinese AI platform named CyberStrikeAI was used to send command scripts.” CyberStrikeAI is a free, public tool. Its own page says it is for testing systems you own or are authorised to test, and it joins more than 100 security tools under an AI agent that plans and runs the steps. Team Cymru, a security research firm, wrote in March 2026 that the code was first published on 8 November 2025 and that it saw 21 servers running it between 20 January and 26 February this year. Read the claim at its real size. Police name the tool in one case of twelve, and say they found “traces of AI” in the frauds. The weak door was the same in every case. What the AI changes is speed and volume: a senior officer said fraudsters now use AI tools “to do this faster and in bulk.” One lender saw 1,131 transfers in three days. A gap that once needed a skilled person for a week can now be worked by software, on many companies at once.

The FBI’s alert on another AI attack agent, and four checks

Five checks for software that sends money

Run these this week.

  • 1. List every rule that trusts a network address. Firewalls, payment gateways, bank connections and partner systems all have them. Test: for each address on a list, ask who else can send from it. If the answer is “any customer of that cloud”, the rule protects nothing on its own.
  • 2. Make every money request prove who sent it. Each request should carry a signature made with a secret key that only your payment system holds, and the receiving side should check it. Test: from an allowed address, send a correct-looking payout request without the signature. It must be refused.
  • 3. Treat a changed bank account as an event, not an edit. When the account on an approved payout differs from the one checked at approval, the payout should stop until a second person confirms it. Test: change the account on an approved test payout and see whether money moves.
  • 4. Put a limit and an alarm on volume. Set how many payouts and how much money can leave in an hour, and who is told when the limit is hit or a new address starts sending. Test: send twenty test payouts in a minute and note who receives an alert, and how soon.
  • 5. Match money out against the people it was meant for, every day. Police say smaller lenders checked weekly or fortnightly and learned of the theft only when customers said the loan never arrived. Test: count the hours between a payout leaving and someone comparing it with the customer’s record.

What we checked on our own

Five apps, 392 doors.

On 17 November 2026 we ran check one on five apps we build and run. They have 392 server entry points between them, the addresses other software can call. We searched their code for every place that reads the caller’s network address and found 11. Seven use it to slow down repeated requests or to count visits. Two write it into the record when a client accepts a proposal. Two use it as an extra rule on top of a password or a sign-in. None lets a request in on its address alone. It is a narrow check and we say so. It reads our own code, not the firewall rules set at our hosting and database providers, and it finds only the usual ways code reads an address. None of the five apps sends a payment; we found no payment or payout library in any of them. So checks two to five above are what we would run with a client whose software does move money, and we have not run them on a system of our own.

AI agents in Dubai, built so each action proves who asked for it

Questions

Questions about IP whitelisting.

  • What is IP whitelisting?

    A rule that accepts requests only from a fixed list of network addresses and refuses the rest. The list is also called an allow-list. It is common on firewalls, payment gateways and bank connections.

  • Is IP whitelisting enough to secure an API?

    No. It narrows where a request can come from, and says nothing about who sent it. If the listed addresses belong to a public cloud, anyone who rents a server there can send from them. Use it as one layer, with a signed request or a login checked on every call.

  • What is CyberStrikeAI?

    A free, public AI tool for security testing, first published in November 2025. An AI agent plans the steps and runs more than 100 standard security tools. Delhi Police told The Indian Express it was used to send commands to one lender in October 2026.

  • Did AI carry out these loan frauds?

    Police named an AI tool in one of the twelve Delhi cases and said they found traces of AI in the frauds. The way in was an ordinary weakness in the lenders’ public apps and a trusted address list. The AI made the work faster and able to run in bulk.

  • How do attackers redirect an approved loan?

    As police describe it: malware in the lender’s public app watches for loans that are approved but not yet paid, the borrower’s bank account is swapped for one the attackers control, and the payout order is sent from an address the system already trusts.

  • Does this matter to a business in the UAE?

    The report is about lenders in India and names no company elsewhere. The weakness is not Indian. Any business whose software sends salaries, supplier payments or refunds through a bank or payment connection that trusts a list of addresses can run the same five checks.

Sources

read 17 November 2026

Check it at the source.

  1. The Indian Express: Hackers are hijacking approved loans. 12 Delhi-NCR firms lose Rs 65 crore (the report of 16 November 2026, with the police account of the method, the twelve cases and the AI tool. We found no police or regulator statement to read beside it)
  2. Team Cymru: Tracking CyberStrikeAI Usage (research of 2 March 2026 on the tool police named: what it is, when it appeared and how many servers ran it)
  3. BRev Digital: AI agents are stealing AI keys to pay for their own attacks (the other attack-agent report of this month, from Amazon)

A question?

Does your software send money?

Tell us what it pays and through which bank or payment service. We will go through the five checks with you and say plainly which door to close first.

Talk to usWhatsApp +971 50 728 5494hello@brevdigital.com