CERT-In Says AI-Assisted Attackers Re-Hit Financial Firms After Remediation
CERT-In says CRAFTY FALCON is re-attacking financial firms with agentic tools, exposing weak remediation, cloud controls, and trusted payment APIs in India.
India's Computer Emergency Response Team issued a critical advisory on November 4, 2026, warning that an AI-assisted threat actor it calls CRAFTY FALCON is conducting an active campaign against financial institutions. CERT-In says successful attacks have reached non-bank lenders, fintech companies, payment providers, aggregators, and cryptocurrency services—including organizations that had already received restricted warnings and attempted remediation.
The campaign matters less because an attacker used a language model than because automation compressed the attack cycle. According to the advisory, operators used several publicly available agentic security frameworks to generate and revise tailored exploits, rotate infrastructure, move through cloud systems, and invoke victims' own payment and disbursement APIs. Some attacks were timed for weekends and holidays, when monitoring coverage was thinner.
CERT-In does not name affected firms, disclose losses, or publish the detailed indicators it shared privately with the financial sector. Its public evidence therefore cannot independently establish the full scope of the campaign or how much of each intrusion was autonomous. But the advisory aligns with separate incident-response reporting from Google's Mandiant team and Kaspersky, both of which say attackers are now using agentic offensive frameworks during real intrusions.
This is a material follow-up to our October analysis of ARTEX and agent memory as forensic evidence. The new warning describes repeat attacks, names a campaign, broadens the tool set, and gives defenders a clearer picture of what changes when intrusion activity can adapt at machine speed.
The financial API is the objective, not merely the entry point
CERT-In's earlier October 9 warning described attackers compromising internet-facing applications, stealing credentials for payment partners and banks, and then calling disbursement APIs directly. That technique bypasses business controls implemented only in a customer-facing application, such as loan approval, identity verification, beneficiary validation, balances, and transaction limits.
The November update says the campaign has escalated. Attackers reverse-engineered application code and modified client packages so they could call back-end APIs without obeying the legitimate application's workflow. They used micro-value transactions—such as a one-rupee transfer—to test whether an unauthorized payment path worked before attempting larger transfers. Merchant-settlement APIs and cryptocurrency wallet keys were also targeted.
This exposes a familiar architectural error: trusting the client to enforce a business rule. A mobile app, web interface, or partner package is observable and modifiable by an attacker. Every consequential rule must be enforced again at the server boundary. A valid credential or request arriving from an expected IP address cannot prove that the surrounding transaction was authorized.
CERT-In says attackers even routed requests through whitelisted shared fetcher pools, including services associated with content acceleration or application review. Blocking an address after an incident is inadequate when unrelated users can cause the same shared service to fetch arbitrary URLs. The advisory recommends identity-based controls such as mutual TLS, signed requests, nonces, timestamps, and scoped keys instead of treating network location as identity.
Agentic tools changed tempo more than technique
The public advisory identifies ARTEX, PentAGI, Cairn, and CyberStrikeAI among the tools used against affected organizations. These systems were presented publicly as penetration-testing or security-research frameworks. An attacker can assign an objective, let agents enumerate paths, generate commands and scripts, analyze results, and change tactics when an attempt fails.
None of the underlying weaknesses is novel. The reported entry points include exposed test systems, vulnerable file-upload and deserialization code, hardcoded credentials, permissive cloud roles, shared access between staging and production, and missing egress controls. What automation changes is how many paths can be tested, how quickly a failed payload can be revised, and how long the campaign can continue without a human operator performing each step.
Google's threat-intelligence and Mandiant teams reported on October 30 that they had observed ARTEX, PentAGI, Cairn, and other harnesses during incident-response engagements since late August. They attributed lateral movement, database exploitation, and preparation for exfiltration to agentic activity that could explore alternatives and troubleshoot errors in seconds or minutes.
Kaspersky separately reported November 2 that attackers used CyberStrikeAI against educational organizations in Latin America and Africa. In its reconstructed case, the implant acted as a conventional remote-control channel while the AI ran on attacker-controlled infrastructure. The final sequence of lateral movement, credential collection, and database export took no more than an hour, according to the company.
Those accounts support the general claim that agentic frameworks are appearing in real compromises. They do not independently verify CERT-In's attribution, affected organizations, or every capability assigned to CRAFTY FALCON. “AI-driven” can blur the distinction between a human directing an adaptive tool and a fully autonomous campaign. Defenders should plan for the faster workflow without assuming the operator has disappeared.
Incomplete remediation becomes an invitation to return
The sharpest finding in the new advisory is that some organizations were attacked again after receiving warnings and undertaking remediation. CERT-In attributes the repeat compromises to incomplete eradication and recommends treating an environment as persistently compromised until forensic verification establishes otherwise.
Rotating one exposed key or patching one server is not enough when an intruder may have created additional identities, changed serverless functions, planted web shells, copied wallet keys, or established management sessions through legitimate cloud services. The recovery record must connect every observed action to the credentials, systems, artifacts, and dependencies it could have affected.
This is where persistent, tamper-resistant operational memory becomes a security control rather than a convenience. An incident team needs a durable timeline of findings, containment decisions, credential rotations, restored assets, unresolved contradictions, and evidence that each remediation step was verified. A summary that silently drops an earlier indicator can recreate the exact blind spot an adaptive attacker will probe.
Controls must operate at machine speed too
CERT-In recommends continuous monitoring of cloud audit logs, rapid containment, default-deny egress, production isolation, short-lived credentials, file-integrity monitoring, server-side transaction validation, and anomaly detection for micro-transactions followed by larger payments. It also recommends detecting bursts of newly generated scripts and parallel attack threads rather than relying only on known malware signatures.
For payment systems, the most important controls sit after authentication: new-beneficiary cooling periods, velocity rules, amount-band anomalies, reconciliation, maker-checker approval, and holds on suspicious disbursements. These mechanisms limit damage even if an attacker has a legitimate token and can reproduce a valid request.
The advisory also tells organizations to monitor unexpected language-model API traffic and agent processes running with permission checks disabled. That is useful telemetry, but banning recognizable tools will not solve the problem. Framework names and models can change. The durable signals are behavior: rapid enumeration, repeated payload mutation, unusual cloud-management actions, cross-environment credential use, and automated payment tests.
CRAFTY FALCON is consequential because the campaign joins three developments that defenders often discuss separately: agentic intrusion tooling, weak cloud boundaries, and payment APIs trusted beyond what their credentials can prove. The practical response is not an “AI firewall.” It is external enforcement, complete incident memory, and controls able to interrupt a valid-looking transaction before an adaptive attacker turns access into money.
Sources
- CERT-In Advisory CIAD-2026-0058: CRAFTY FALCON campaign and recommended actions
- CERT-In: October warning on payment and disbursement API attacks
- Google Threat Intelligence Group and Mandiant: agentic frameworks observed in intrusions
- Kaspersky: CyberStrikeAI activity observed in compromised networks
- Reuters: Indian authorities warn financial firms about AI-assisted attacks
Your AI remembers everything. Everywhere.
Unimatrix gives you a shared, durable memory layer across Claude Desktop, Cursor, ChatGPT, and Gemini. Setup in 2 minutes. Free and paid plans available.
Looking for developer resources?
Browse our catalog of 500+ tested AI prompt profiles covering DevOps, data modeling, agent behaviors, and API wrappers. Have a prompt to share? Submit your own for review by our librarian to be featured. Completely free, no registration required. Browse prompt libraries →
Keep reading
Autoware 2.0 integrates learned driving planners, but its proposed runtime safety selector does not yet operate in Autoware’s present open-source release.
A Nature paper demonstrates a designed molecular spin–photon interface, but scalable quantum chips still require integration, entanglement, and lower loss.
NVIDIA’s open-weight Nemotron 3.5 Transcribe turns recorded speech into structured agent memory, raising serious provenance, privacy, and consent risks.