Blog da Zscaler
Receba as últimas atualizações do blog da Zscaler na sua caixa de entrada
NYDFS Part 500: Your AI Evidence Must Exist by December 31
The Certification That Outlasts Your Job
TL;DR: 2026 is the first full year New York Department of Financial Services (NYDFS) requires an inventory of every information system, which NYDFS guidance says includes AI, and the written procedures that produce and maintain it. In April, two named executives must certify compliance, and the records supporting the certification must be kept for five years.
In 2031, someone asks what evidence supported your 2027 certification.
An examiner, or a plaintiff's attorney, wants to know how the AI systems your organization used in 2026 were inventoried, what access they had, and where the records are. The CISO who signed no longer has a login to any system that could answer. The people who might remember have moved on. And if the evidence was never collected, no one can produce it.
That certification is due on April 15, 2027. Two people sign it: your highest-ranking executive, usually the CEO, and your CISO. It covers calendar year 2026, and the strongest evidence is what was created during 2026.
If that evidence doesn't exist at your institution yet, you have only weeks to start collecting it.
2026 Is the First Full Year Under the New Inventory Rule
Section 500.13(a) has applied since November 1, 2025. It requires two things, and they are easy to read as one. It requires written policies and procedures designed to produce and maintain a complete, accurate, and documented asset inventory covering every information system, not just those that are material or hold customer data. The procedures must include a method to track, as applicable, five attributes: owner, location, classification or sensitivity, support expiration date, and recovery time objective. It also requires that the inventory itself be maintained in accordance with those policies, with a stated frequency for updating and validating it.
So there are two ways to fall short. A list with no written policy behind it does not meet the first requirement. A policy that never produced a list does not meet the second. There is also a third position, and it is the one worth planning against. An institution that assembles the list in March 2027 produces it without any record of the process that was supposed to produce it during 2026. That record is what has to support the certification.
Where AI Fits in the Inventory
The regulation does not name AI. NYDFS's October 2024 guidance does.
Most of 2026 is already behind you. The evidence for it is either being collected now, or it will never exist.
The certification asserts material compliance across your entire cybersecurity program, including the AI you already use: copilots, agents with access to customer data, and AI built into third-party vendors' services. At many institutions, AI arrived faster than the inventory did.
What Enforcement Has Actually Done
In the certification actions NYDFS has published through September 2026, the Department has charged institutions rather than the individuals who signed.
Results have been mixed. In an April 2021 consent order, NYDFS found that an NYDFS-licensed firm had falsely certified compliance with the regulation for calendar year 2018 and imposed a $3 million penalty. And in a 2021 pleading against a title insurer, the Department noted that the company's CISO had signed the certification. It charged the company. Those certification charges were not in the consent order that settled the matter two years later.
So the Department points at the signature and charges the institution. But the names are personal, and the people who signed are asked the questions that follow. The institution may pay the legal fees; the signers still spend the time describing a decision made years earlier at a different job.
What to Do Before December 31
Three steps determine whether your records will support your certification:
Find the AI already in use. That includes approved tools, unapproved ones, and AI built into vendor products.
Collect the evidence as it happens. Records reconstructed in March are weaker than records created in October.
Decide what to sign based on what the evidence shows. Don't let the deadline make that decision.
In September 2026, NYDFS said the asset inventory should be the foundational input to your risk assessment, so an incomplete AI inventory creates problems for more than one Part 500 obligation.
When the evidence exists, the certification conversation changes. The signers review a documented year instead of reconstructing one from memory. Our white paper on proving control over AI in financial services sets out what that record should include.
How Zscaler Helps You Build the Record
Because the Zscaler Zero Trust Exchange platform sits in line between your users and the applications they reach, AI use is visible as it happens, not reconstructed afterward. Zscaler Security for AI uses that position to produce evidence for each of the three steps.
Find the AI already in use. Zscaler recommends a 30-day first phase: discover the AI applications in use, classify each one as sanctioned, unsanctioned, or unreviewed, and start inspecting what goes into them. Discovery happens in line, in the traffic Zscaler already inspects: which AI applications employees use, approved or not. Zscaler extends that to the AI built into approved SaaS applications, and to the models and agents an institution builds itself, which are the systems an inventory most often misses.
Collect the evidence as it happens. Every allow, warn, block, and isolate decision on AI traffic produces a log entry. Because the Zero Trust Exchange brokers each connection directly between the user and the application without placing them on the network, these entries capture AI use with identity, device, and policy context that can feed your inventory, not a retroactive reconstruction from network flow data.
Decide what to sign based on what the record shows. By April, the signers can read a year of activity: which AI was used, by whom, with what data, and which policy actions were taken. They won't have to rely on memory when someone asks.
See Where Most Organizations Start
In its 2026 Frontier AI Readiness Report, Zscaler ThreatLabz assessed hundreds of large enterprises, and financial services was the largest group. Of those organizations, 91% had AI endpoints exposed to the public internet, and every organization assessed lacked identity-based authentication for its AI. Read the report
Then find out where your organization stands. Talk with a Zscaler expert about where AI risk may exist in your environment and what to address first. Talk with an expert
The signature takes a few minutes. The records have to survive for five years. When someone asks in 2031 what evidence supported your certification, the answer must come from records collected in 2026, not from memory.
FAQs
Before 2023, the certification form could be signed by the board chair or a senior officer. The 2023 amendments moved the signature into the regulation itself and replaced that choice with two named individuals signing together: the highest-ranking executive and the CISO. Where an institution has no CISO, Section 500.17(b)(2) substitutes the senior officer responsible for the cybersecurity program. The first certification under the amended rule fell due in April 2024.
NYDFS has written about AI after each of the last three certification cycles. In October 2024, it explained how Part 500 applies to AI. In October 2025, it suggested vendor contracts consider addressing AI, including whether vendors may train AI on your data. In May 2026, it issued an advisory on frontier AI models addressed only to CISOs, the officer who co-signs the certification. None of the letters created new requirements. Each one made it harder to say AI was out of scope.
Each certification must be supported by data and documentation sufficient to show that you complied. When the 2023 amendments changed the attestation from "is in compliance" to "materially complied," NYDFS still didn't define "material," though its FAQs give examples. Your organization makes that judgment for itself, so if the decision is questioned, the signers must explain their judgment.
Part 500 requires the records supporting each certification to be kept for five years. Since 2023, that includes the areas found to need improvement, the remediation work, and the plans and timelines for it. Records for the April 2027 certification must be available until 2032. Five years is the minimum. It doesn't limit when questions can be asked.
Part 500 offers a second option. An acknowledgment of noncompliance names the requirements you did not meet, describes the gap, and provides a remediation timeline or confirms remediation is complete. The same two people sign it, and on its face, it offers no protection. But unlike a certification of compliance, it can be accurate even when the records are incomplete.
Esta postagem foi útil??
Aviso legal: este post no blog foi criado pela Zscaler apenas para fins informativos e é fornecido "no estado em que se encontra", sem quaisquer garantias de exatidão, integridade ou confiabilidade. A Zscaler não se responsabiliza por quaisquer erros, omissões ou por quaisquer ações tomadas com base nas informações fornecidas. Quaisquer sites ou recursos de terceiros vinculados neste post são fornecidos apenas para sua conveniência, e a Zscaler não se responsabiliza por seu conteúdo ou práticas. Todo o conteúdo está sujeito a alterações sem aviso prévio. Ao acessar este blog, você concorda com estes termos e reconhece que é de sua exclusiva responsabilidade verificar e utilizar as informações conforme apropriado para suas necessidades.
Receba as últimas atualizações do blog da Zscaler na sua caixa de entrada
Ao enviar o formulário, você concorda com nossa política de privacidade.


