Skip to content
RESEARCH INDEX BREACHROAD / INTELLIGENCE NOTE

CRAFTY FALCON: AI attacks on payment APIs

CERT-In warns of an active AI-assisted campaign against banks and fintechs. We analyse UAT, cloud control planes, payment APIs and defence.

PUBLIC RESEARCH
AUTHOR
/ CEO of Breachroad · OSCP · PNPT
PUBLISHED
4 November 2026
READING TIME
16 min read
TOPIC
AI Security
CRAFTY FALCON: AI attacks on payment APIs

CERT-In issued critical advisory CIAD-2026-0058 on 4 November 2026 for an active campaign against the financial sector. CSIRT-Fin named the operator CRAFTY FALCON and describes a financially motivated, AI-assisted adversary that automates reconnaissance, generates script variants and invokes victims’ own payment and disbursement APIs. Targets include banks, fintechs, digital lenders, payment aggregators and virtual-asset services.

The official CERT-In advisory updates a warning from 9 October. CERT-In says attacks remain active and that some previously warned organisations were targeted again after incomplete remediation. The operator selected holidays, weekends and reduced-monitoring windows to delay detection of unauthorised transfers.

What CERT-In confirms

The advisory confirms:

  • successful incidents involving NBFCs, fintechs and other BFSI entities;
  • repeated targeting of some previously notified organisations;
  • use of ARTEX, PentAGI, Cairn and CyberStrikeAI;
  • automated generation and refinement of many script variants;
  • infrastructure rotation over hours or days;
  • entry through public UAT, staging and test environments;
  • movement to production through shared credentials, tokens and network paths;
  • abuse of AWS Systems Manager, Lambda and other legitimate management services;
  • reverse engineering of application code to build tailored exploits;
  • payment and disbursement API calls from modified application packages;
  • an ongoing campaign requiring immediate action.

CERT-In characterises some tools as autonomous and machine-speed. This does not prove that no human participates or that every stage of every attempt ran without supervision. For defenders, the important point is that the volume of attempts, variants and decisions can exceed manual SOC analysis speed.

The attack chain

The operation does not begin at the payment API. Its complete model spans several layers:

  1. reconnaissance of public hosts, repositories, mobile applications and dependencies;
  2. entry through a less protected non-production environment;
  3. context collection from code, configuration, secrets and application traffic;
  4. movement into production through shared identity or flat networking;
  5. use of legitimate cloud control channels for commands and file delivery;
  6. reproduction of the client protocol and generation of tailored requests;
  7. financial API invocation that resembles a legitimate application;
  8. infrastructure rotation and retargeting after partial mitigation.

A central defensive error is assuming a request from the official application package is trusted. If an attacker can reproduce the format, steal a token or alter the client, a mobile-side control does not protect a transaction. The backend must enforce authorization and risk rules for every operation.

Why UAT becomes a path to money

Test environments often expose verbose logging, debug features, sample accounts, a weaker WAF and longer-lived secrets. Yet they also need production-like data and integration. A “non-production” host can therefore contain a production token, repository access or a network route to settlement systems.

Compare UAT and production for:

  • shared service accounts and keys;
  • common network zones;
  • copied databases and personal data;
  • real queues, webhooks and APIs;
  • deployment-pipeline permissions;
  • access to Systems Manager, Lambda and administrative functions;
  • differences in EDR, WAF, logging and retention.

Domain-name separation is not a security boundary. If UAT can retrieve a production secret or execute a production command, it belongs to the same trust zone.

Detection that survives IOC rotation

CERT-In warns that infrastructure changes quickly, so blocklists alone are insufficient. More durable signals include:

  • many functionally similar scripts created in a short period;
  • attempt series differing in syntax, encoding or parameters;
  • new SSM or Lambda activity from hosts that do not normally use it;
  • UAT traffic to a production database, queue or console;
  • an application package with a changed digest calling valid endpoints;
  • a new device or ASN performing many otherwise valid transfers;
  • a spike in schema reads and validation errors followed by success;
  • transfers in low-monitoring windows or after a limit change;
  • the same token used from different geographies or TLS stacks;
  • interaction with honeypots, canary credentials and decoy consoles.

Measure the sequence, not only a transaction: reconnaissance → error → adaptation → success → disbursement. AI can change payloads rapidly, while the business objective and stage relationships remain more stable.

Controls for payment APIs

The backend must treat the client as untrusted. A minimum design includes:

  • strong service and user authentication;
  • short-lived, audience-bound tokens;
  • nonce, idempotency keys and replay protection;
  • signatures over critical transaction fields;
  • value, velocity and beneficiary limits;
  • step-up authentication for new devices or recipients;
  • separation of creation, approval and disbursement;
  • behavioural rules independent of request validity;
  • an emergency stop that does not disable the entire organisation;
  • a complete audit trail for decisions and configuration changes.

Validate these mechanisms through a REST, GraphQL and OAuth API penetration test and web/API penetration testing.

Campaign response plan

  1. inventory public UAT, staging, test and developer consoles;
  2. remove shared secrets and production paths;
  3. review SSM, Lambda and control-plane activity;
  4. hunt rapid script variants and modified application packages;
  5. isolate hosts with confirmed execution;
  6. stop high-risk payouts and preserve transaction chronology;
  7. rotate tokens after removing active access;
  8. validate beneficiaries, limits and rule changes;
  9. report to competent authorities and financial partners;
  10. retest against the same objective, not only a known IOC.

When an organisation is hit again, determine why remediation was incomplete: a surviving identity, token, network trust, backdoor or unchanged API logic. Use an AI incident-response playbook and structured Cyber Threat Intelligence lifecycle.

Sources versus Breachroad conclusions

CERT-In confirms sector scope, active attacks, named frameworks, UAT entry, cloud management channels, modified packages and payment APIs. Detailed infrastructure indicators are distributed to BFSI entities under restricted handling.

Our backend-authorization model, behavioural correlations and response order are Breachroad conclusions. AI red teaming can test agents and automation boundaries, while secure AI implementation should cover identities, tools and scope control. The main lesson is that attack speed increases the value of fundamentals: segmentation, short-lived secrets and automated detection.

Payment APIs: control outcomes, not application trust

A payment API should re-evaluate amount, beneficiary, limit, order state and identity role even when a “trusted” backend submits the request. Transaction signatures must bind every material field rather than only a session identifier. Changing an account after approval requires new authorisation. A modified application package then cannot freely transform a legitimate workflow.

High-risk operations require an out-of-band control: dual approval, delay, daily cap or confirmation by another identity. The same panel being protected should not control every rule. Alerts should reach an independent system, and responders need to freeze a token or beneficiary without stopping the entire platform.

UAT must not be a shortcut to production

Test environments commonly retain integration copies, shared accounts or cloud roles that can reach production. Separate tenants, KMS keys, secrets, CI and service identities. Test data should be synthetic and access to a live payment API denied by default.

Monitor AWS SSM, Lambda and other management channels used by UAT identities. A legitimate deployment has a change ID, expected artefact, signature and window. A command lacking those elements should stop for review. Export cloud audit records to an account the application team cannot alter.

Detecting package modification

Every artefact needs a digest, signature, SBOM and provenance from commit to production. Deployment systems should reject packages built outside the approved pipeline. Monitor executable-file differences on hosts, but compare them with the release manifest rather than relying on a generic scanner.

At runtime, alert on new endpoints, domains, libraries and payment-API calls. Join the application request to the transaction seen by the provider. A discrepancy in beneficiary, amount, time or device identifier is stronger than one attacker-infrastructure IOC.

A financial-sector exercise

A tabletop can start on Friday evening: UAT creates plausible SSM sessions while production sends transactions to new beneficiaries. Participants decide how to disable identity, stop payments, preserve logs and communicate with customers. Later, reveal a shared secret and modified artefact.

Measure time to freeze financial impact, not only time to name malware. Can the organisation continue essential payments manually? Does the weekend contact list work? Who owns the decision? Cybersecurity training should rehearse this conflict between availability and integrity.

Improvement metrics

Quarterly, measure the share of short-lived production identities, UAT-to-production shared secrets, signed artefacts, token-revocation time and transaction value that can be stopped before settlement. Every metric needs an owner. Success is not an AI detector purchase; it is breaking the path before modified code creates an irreversible outcome.

A financial-release checklist

Before deployment, confirm artefact provenance, digest, SBOM, payment-API test and absence of UAT-shared secrets. During the window, observe SSM, Lambda, CI operator, beneficiaries and amounts. The first production cohort should have a low limit and immediate rollback.

Afterwards, match transactions to the approved change, review new connections and preserve evidence. Any manual package edit requires a fresh build and signature; changing files directly on a server destroys provenance. Release control, rather than recognising one AI framework name, remains effective when adversary tools change.

Payment owners should formally accept the first cohort’s risk cap and receive the deployment report. Otherwise rollout speed remains a technical decision despite its financial consequence.

Keep that approval with the artefact, test results and canary observations in one evidence record.

SHARE / COPY