Back to News CERT-In has released a comprehensive advisory against AI enabled hacking and business process flow compromise.
CERT-In, 20 November 2026 · Severity: Critical.
Attackers are changing the records behind payments and letting legitimate approvals pay out.
Separately, AI-driven agentic offensive frameworks (AOFs) are being used in attacks across sectors.
Advisory Brief #
1. Abuse of business payment workflows
Unauthorised transactions are no longer limited to banks and financial firms.
They also hit technology service providers supporting government-linked payment processes, and platforms handling vendor or partner payouts.
How attackers did it:
- Misused business applications and provider-operated workflows, not just payment APIs.
- Altered business records or approval relationships, then used legitimate payment processes.
- Used trusted systems and normal approval steps, so nothing looked suspicious.
- Exploited weak independent verification, reconciliation and review of record changes.
Crux: a payment can pass every application and approval check even when the records behind it were changed without authorisation.
Verify the record and the authority for a payment, not just the payment request.
2. AOF-enabled intrusions across sectors
CERT-In's threat hunting found AOFs used against enterprises, technology and service providers, government organisations, and academic and research institutions.
Techniques seen:
- Automated reconnaissance of exposed systems, services and attack paths.
- Fast building or modification of exploits for known flaws, including by analysing software updates.
- Multiple AI agents running reconnaissance, privilege discovery and other tasks in parallel.
- Attack steps that adapt to errors, blocks or responses from the target.
- Valid credentials, legitimate admin functions, trust relationships and external infrastructure used to go deeper.
The crux: AOFs need less manual effort and shorten the time between reconnaissance and the next attack stage.
What CERT-In asks organisations to do
How to read it. P0 = do this week. P1 = do this month. P2 = plan for this quarter.
The tag at the end shows where the item comes from: [20 Nov] is the advisory above; [0046] [0049] [0058] are the earlier CERT-In advisories it refers to; [Blueprint] is CISG-2026-02; [Directions 2022] is the CERT-In rulebook from April 2022; [Good practice] is sensible advice, not a CERT-In requirement.
1. Know what can move your money
- P0 We have one list of every person, system, vendor, platform and integration that can change who gets paid, approve a payment or release funds. This includes SaaS finance tools and AI tools. [20 Nov]
- P0 Every account on that list is actually needed. Anything extra is removed. [20 Nov]
- P1 If we handle government-linked payments, vendor payouts or partner settlements, we protect them as carefully as a bank would, even though we are not one. [20 Nov]
2. Payment records and approvals (the heart of this advisory)
- P0 Nobody can change a vendor, beneficiary or bank detail alone. One person makes the change, another checks it, and the system records who, what, when and from where. No direct database edits, no bulk uploads without approval. [20 Nov]
- P0 Any change to bank details is confirmed by calling the vendor on a number we already had on file, never a number from the email or invoice asking for the change. The call is recorded. [Good practice]
- P0 Changes to who approves payments, or to payout limits, need the same two-person check as bank details. [20 Nov]
- P0 New payees wait before they can receive large amounts, and get lower limits until verified. Many banks already use a 24-hour window; 24–48 hours is a sensible starting point. [0058; hours are good practice]
- P0 No emergency override skips the approval chain. Any exception needs a second approver, expires, and is reviewed within 24 hours. [Good practice]
- P0 Before paying, finance matches the invoice to the purchase order, proof of delivery and the vendor record. An invoice alone is not proof. [Good practice]
- P1 High-value and first-time payments need two approvals, and the system enforces it, not just a policy document. [Good practice]
- P1 For risky changes, at least one approval happens outside the system that holds the payment. Several approvals inside one compromised system can all be fooled at once. [20 Nov]
- P1 Approved payment batches are matched against bank statements and payout reports the same day. A mismatch puts payments on hold. [Good practice]
- P1 Old, unused and duplicate vendors are removed, including duplicate PANs or bank accounts. [Good practice]
3. If a service provider runs your payments or holds your records
- P0 We know exactly which provider accounts and integrations can change our payment records, approve payments or see sensitive data, and each one is justified. [20 Nov]
- P0 We regularly compare the provider's payment records with our own master copy, kept somewhere the provider cannot edit. [20 Nov]
- P1 Our contract covers: how fast they must tell us about a breach, our access to the relevant logs, help with investigation, and our right to affected payments or integrations. [20 Nov]
- P1 The contract also says they will warn us promptly if a breach at another client could affect us. [20 Nov]
4. If you are a service provider, ITES firm or payout platform
- P0 Access given for one client can never reach another client's systems or data. [20 Nov]
- P0 Admin and control-panel accounts are few, named, protected with strong sign-in and reviewed often. [20 Nov]
- P1 We watch changes to payment records together with admin logins, approvals and the payments that follow, and investigate odd sequences even when each step came from an authorised account. [20 Nov]
- P1 If we suspect a breach, we check the affected records against an independent source and ask: what other clients, integrations or systems could this account reach? [20 Nov]
- P1 Vendors using our portal see only their own invoices and payments, and cannot change approval workflows. [Good practice]
5. Payment and disbursement APIs (Critical for NBFCs)
- P0 Every transaction is checked on our servers: amount, payee, currency and limits. We never trust the app or partner system alone. [0058]
- P0 Payment requests are digitally signed and can't be replayed. Unsigned or repeated requests are rejected. [0058]
- P0 Test, staging and live systems are fully separate: no shared passwords, keys, databases or network paths. Any test-to-live access raises an alert. [0058]
- P0 No passwords, keys or tokens sit in code, config files or public repositories. Exposed ones are changed and moved to a secure vault. [0049]
- P1 We don't rely on "trusted IP address" lists alone. Each partner gets its own narrow, short-lived credentials. [0058]
- P1 Our apps detect tampered or repackaged copies, and partner connections are mutually authenticated. [0058]
- P1 File-upload and data-handling flaws in internet-facing PHP and Node.js apps are patched first. [0058]
- P1 Web folders can't be changed at run time, and we're alerted to any unexpected file (a sign of a planted web shell). [Good practice]
- P2 Rate limits block password-guessing and floods of requests from one key or partner. [Good practice]
- P2 If a partner routes traffic through a shared service, we log the real origin and key, not just the shared IP. [0058]
6. Vendors and partners
- P0 New vendors are verified before activation: GST, PAN, bank account (for example a small test deposit or bank confirmation) and authorised signatory. [Good practice]
- P0 When a vendor or partner relationship ends, their keys, portal access and payout rights are removed the same day, and we check it worked. [Good practice]
- P2 Once a year we review access, keys and incident history for our top vendors by payout value. [Good practice]
- P2 Nobody pays company money through a personal UPI, wallet or unapproved payout tool. [Good practice]
7. Logins and access
- P0 Every login to finance, admin, payment and cloud systems uses multi-factor sign-in. No exceptions for senior leaders, vendors or service accounts. Phishing-resistant methods where possible. [20 Nov]
- P0 Admin access is granted only when needed, for a limited time, approved and recorded. [Good practice]
- P1 System and service accounts use narrow, short-lived keys, and use outside normal hours raises an alert. [0049]
- P1 Access is reviewed when people change roles and removed within 24 hours of leaving. [Good practice]
- P1 Cloud consoles are open only to named admins on managed devices. [Good practice]
- P2 No shared logins for finance or admin. Sessions time out. Every month we review who can create vendors, approve payments and release funds. [Good practice]
8. AI agents and AI-driven attack tools
- P0 We keep a list of every AI agent, copilot, automation bot and AI coding tool that can read data, send messages or trigger payments, including ones staff set up themselves. [0049]
- P0 No AI agent can create vendors, approve payments or move money without a human approving it. [Good practice]
- P0 Agents can reach only approved systems. Everything else is blocked by default at the network level, not just by instructions to the agent. [0049]
- P0 Each agent has its own identity and its own narrow, short-lived keys. Agents never get banking passwords, OTPs or anyone's personal logins. [0049]
- P0 We have a tested way to stop all agents immediately. [Good practice]
- P1 Agents ignore instructions hidden in emails, web pages, files or calendar invites, and can use only approved tools. [Good practice]
- P1 Everything an agent does is logged, and agents cannot edit or delete their own logs. [Good practice]
- P1 AI-written code is reviewed, scanned and tested before it touches payment systems. [Good practice]
- P1 AI tools under trial run only in isolated setups with dummy data, never on live systems. [0049]
9. Internet exposure, cloud and devices
- P0 We know every system exposed to the internet, including old test sites, forgotten APIs and admin panels. Admin panels are not open to the internet. [20 Nov]
- P0 Actively exploited flaws on internet-facing systems are fixed within 12 hours where feasible. [Blueprint]
- P0 Cloud admin activity is watched in real time, especially new keys and policy changes. [0058]
- P1 Servers and payment systems can only connect out to places they need. [Good practice]
- P1 Payment systems sit on their own network segment, away from office and guest networks. [Good practice]
- P2 Email is protected against spoofing (SPF, DKIM, DMARC), and fake "finance" or "vendor" emails are flagged. [Good practice]
- P2 Payments are approved only from managed, updated, encrypted devices. Finance backups are isolated and test-restored every quarter. [Good practice]
10. Watching and detecting
-
P0 Someone watches payment systems, APIs, cloud and logins 24x7, including weekends and holidays. [0058]
-
P0 At night, on weekends and on holidays, payment alerts get stricter, not looser. [Good practice]
-
P0 We get an alert when a small "test" payment to a new payee is followed by a large one. [0058]
-
P1 Any change to payees, approvers or payout limits is a high-priority alert that someone reviews immediately. [20 Nov]
-
P1 Login, device, app, network, cloud, API and agent logs flow into one place where they can be connected. [20 Nov]
-
P1 Each week we look for signs of trouble: new payees followed by big payments, unusual logins and admin actions, large data transfers out, and test-to-live access. [20 Nov]
-
P2 Once a quarter we rehearse one payment-fraud and one AI-misuse scenario. [Good practice]
11. When something goes wrong
- P0 We have a written, practised playbook: freeze payouts, call the bank, preserve logs, notify CERT-In. [Good practice]
- P0 If money has moved, we call the bank's fraud team on its official number at once, ask for a hold or recall, and confirm in writing the same day. [Good practice]
- P0 For financial fraud, we call 1930 and file on cybercrime.gov.in. [I4C]
- P0 One named person owns reporting to CERT-In within 6 hours of noticing an incident. Reporting doesn't wait for legal sign-off. [Directions 2022]
- P1 Logs, signed requests, agent logs and payment files are preserved before anything is cleaned up. [Directions 2022]
- P1 We assume the attacker may still be inside, and check for leftover access, new keys and new vendors before closing the incident. [0058]
- P1 Banks and financial firms know how CERT-In reporting lines up with CSIRT-Fin and RBI reporting, with timelines and contacts. [0058]
- P1 Finance, security, legal and communications know who is called, in what order, and who decides what. [Good practice]
12. People and yearly checks
- P0 Finance and vendor teams are trained on bank-detail change scams, invoice redirection, approval bypass and AI misuse, with real examples. [Good practice]
- P1 Every quarter we run a fake "change our bank details" or "urgent CEO payment" test. [Good practice]
- P1 Developers of payment APIs are trained on server-side checks, request signing and keeping secrets out of code. [Good practice]
- P2 Annual security testing covers payment APIs, the vendor portal and AI agent connections. [Good practice]
- P2 Once a year we confirm CERT-In Directions compliance: 6-hour reporting, logs kept for 180 days in India, and a named contact person registered with CERT-In. Cloud, data centre, VPN and crypto businesses also check their 5-year record-keeping duties. [Directions 2022]
- P2 After every incident, we update this checklist, playbooks and controls. Two weeks is a reasonable deadline. [Good practice]