All postsCybersecurity

CERT-In advisory on AI cyber attacks: CRAFTY FALCON

APJ-RAISE team · · 6 min read

The CERT-In advisory on AI cyber attacks, CIAD-2026-0058, warns that an AI-assisted actor called CRAFTY FALCON is moving money out of Indian financial firms by calling their own payment and disbursement APIs. CERT-In issued it on 4 November 2026, rated it Critical, and updated it on 18 November 2026.

A phone screen showing a bank alert about unusual activity in an account
Photo: RDNE Stock project, Pexels

Most summaries copy the advisory's list of techniques. A list is hard to defend against, because you cannot see where one step hands over to the next. So this post asks one question: what does an AI-assisted fintech attack look like, step by step, and where does each step leave a trace? Everything comes from CERT-In's public text. The order of steps is our reading, since CERT-In does not give a fixed sequence.

CERT-In advisory CIAD-2026-0058 in brief

The name CRAFTY FALCON comes from CSIRT-Fin, the finance-sector response team under CERT-In. The targets are NBFCs, digital lenders, fintechs, payment aggregators, payment service providers and crypto exchanges. CERT-In says the threat is active and ongoing.

Three details matter more than the rest. The actor returns to firms it already hit, including ones that applied earlier restricted advisories. It likes holidays, weekends and other quiet hours. And it does not need to break a bank's core systems, because the company's own APIs already move money.

The public page does not say what changed in the 18 November update. A restricted version with indicators of compromise and detection rules goes only to BFSI entities (banking, financial services and insurance).

Which AI tools does CRAFTY FALCON use?

CERT-In names four tools. ARTEX is a family of LLM-based tools; Microsoft describes it as an open-source, multi-agent pentest framework that can go from scouting to data theft with little human input. PentAGI is an autonomous multi-agent pentest system. CERT-In calls Cairn an attack search engine and CyberStrikeAI a framework.

All four started life as security testing tools. CERT-In points out that none of them checks whether you are allowed to test a target, so they attack as well as they test. ARTEX also turned up in the Storm-0940 Microsoft 365 campaign and in attacks on South Korean financial firms reported by Wiz. Our ARTEX explainer covers the tool itself.

How the CRAFTY FALCON attack works, step by step

To make the chain concrete, picture a made-up lending app, "Lender X". Nothing below is about a real company. Each step is a technique from the advisory, placed in the order we think makes sense, with the trace it would leave in Lender X's logs.

  1. 1Find a test server. The actor locates a UAT or staging server open to the internet. Trace: unknown IPs scanning or logging in to a non-production host.
  2. 2Cross into production. It reuses credentials, tokens or network paths that test and live systems share. Trace: a login to production from a test machine, which should never happen.
  3. 3Take the app apart. It reverse-engineers the Lender X mobile or web app and calls the APIs with a modified app package. Trace: requests that fail app attestation, or come from an unknown app build.
  4. 4Hide behind trusted IPs. It routes traffic through third-party "fetcher" services whose addresses Lender X already allow-lists. Trace: a trusted IP suddenly calling payment endpoints it never used before.
  5. 5Test the pipe with Rs. 1. A tiny transfer to a new payee checks that money really flows. Trace: a Rs. 1 payout to a fresh account.
  6. 6Drain. It abuses disbursement APIs, or a payment aggregator's merchant settlement API, to move large sums. Trace: a big transfer to that same new payee soon after the Rs. 1 test.
  7. 7Stay inside. It uses cloud tools such as AWS Systems Manager and Lambda to run commands and drop files. Trace: new remote commands, new Lambda code, new roles or access keys.

CERT-In also lists quick use of fresh bugs, especially file-upload and deserialization flaws that plant web shells in PHP and Node.js apps. It mentions theft of crypto wallet keys, and bursts of many similar custom scripts, a pattern that hints at AI-written code.

Look at the trace column again. Every step is visible to someone, but each trace lands in a different log, owned by a different team: network, identity, app security, fraud and cloud. AI speed works best in the gaps between those teams, and a long weekend widens the gaps.

The Rs. 1 rule, as a worked example

The Rs. 1 test is the easiest step to catch, and CERT-In tells firms to hold or review tiny transfers to new payees that are followed by large ones. A simple rule that matches that advice could read like this:

  • Trigger: a payee added in the last 24 hours receives a transfer of Rs. 10 or less.
  • Then: any transfer above a set limit to that payee in the next few hours is held for a human check.
  • Extra weight: raise the risk score on weekends and public holidays.

The thresholds here are our example, not CERT-In numbers. Tune them with your own fraud data. The point is that an AI-assisted attacker still has to check its work before the big move, and that check is visible.

How is CIAD-2026-0058 linked to CIAD-2026-0049?

CIAD-2026-0049, issued on 15 October 2026 and also Critical, is CERT-In's wider warning on AI agents as attack vectors. It describes two problems. AI model providers told CERT-In that their models, during research and testing, reached live systems of Indian organisations using exposed credentials. CERT-In added that the victims did not spot this access by themselves.

The second problem is a money-driven campaign using public open-source offensive AI tools. CERT-In says these tools shrink the time from first access to damage from days to a few hours. Victims who only partly fixed things were hit again within days.

CIAD-2026-0058 also builds on CIAD-2026-0046 of 9 October 2026, on abuse of payment and disbursement APIs. By our count, that makes three linked advisories in 26 days (9 October to 4 November). Few outlets point out that pace, and it tells you CERT-In sees this as one fast-moving campaign, not three stories. The shared weak spot is credentials that work where they should not, which our post on AI agent identity unpacks.

What CERT-In asks fintech teams to do

CERT-In's countermeasures map neatly onto the steps above:

  • Separate UAT: no shared credentials, tokens or network paths between test and production. Alert on any test-to-production access.
  • Drop IP trust: replace IP allow-lists for third-party fetchers with identity checks such as mutual TLS, signed requests or scoped API keys.
  • Distrust the app: validate every transaction on the server. Sign requests with a nonce and timestamp, and use app attestation.
  • Slow new payees: velocity rules and cooling periods, plus the Rs. 1 check.
  • Watch the cloud: real-time alerts on remote commands, serverless code changes and new roles, users or keys.
  • No secrets in code: scan repos and CI/CD, and use short-lived, least-privilege tokens.
  • Patch fast: short deadlines for internet-facing apps, starting with file-upload and deserialization flaws.
  • Staff the weekend: 24x7 monitoring, with more people and lower alert thresholds on holidays.
  • Assume they stayed: after an incident, clean up fully, verify with forensics and keep hunting.

Report suspicious activity to CERT-In at [email protected] with logs. CERT-In asks that fraudulent transactions be reported in bulk to I4C (the Indian Cybercrime Coordination Centre) through its CFCFRMS portal.

If you are a student building a payment project, the same two habits protect you: never share one API key between test and live, and never put payment keys in app code or a public repo. CERT-In also wants every pentest to have written approval, a scope and a time window, and that rule applies to your AI security tools too. To learn how these agents work, start with what AI agents are.

Questions people ask

What is CERT-In advisory CIAD-2026-0049?
CIAD-2026-0049, issued on 15 October 2026, is CERT-In's Critical warning on AI agents as attack vectors. It says AI agents under evaluation reached live Indian systems using exposed credentials, and that attackers use open-source AI offensive tools.
What is CERT-In advisory CIAD-2026-0046?
CIAD-2026-0046, issued on 9 October 2026, covers abuse of payment and disbursement APIs. CIAD-2026-0058 on CRAFTY FALCON builds on it.
What does the CERT-In advisory on AI security say?
CIAD-2026-0058 says the actor CRAFTY FALCON uses AI tools such as ARTEX, PentAGI, Cairn and CyberStrikeAI to abuse payment APIs at Indian NBFCs, fintechs and crypto exchanges, and lists countermeasures.

Sources

  1. 1.CERT-In. CIAD-2026-0058: Escalation of AI-Driven Attacks on BFSI Entities by Threat Actor "CRAFTY FALCON". 4 November 2026, updated 18 November 2026.
  2. 2.CERT-In. CIAD-2026-0049: AI Agents as Attack Vectors: Unauthorized Access by Agents under Evaluation and Misuse of Open-Source Agentic AI Tools Targeting Indian ICT Assets. 15 October 2026, updated 18 November 2026.
  3. 3.CERT-In. Advisories list, 2026.
  4. 4.Microsoft Threat Intelligence. AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise. 21 October 2026.
  5. 5.Wiz. The Rise of Agentic Attacks. 29 October 2026.
All posts