September 11, 2026 · KeyCaliber

NYDFS's Risk-Assessment Guidance Asks Four Questions Most Firms Can't Answer

On September 10, 2026, NYDFS published guidance on what a Part 500 risk assessment has to contain — and a list of the failures it keeps finding. Four of its demands come back to the same thing: an asset inventory good enough to reason from.

Asset Visibility Compliance Shadow AI Third-Party Risk

On September 10, 2026, the New York Department of Financial Services issued industry guidance on cybersecurity risk assessments. It creates no new rule. It explains what DFS expects a Part 500 risk assessment to contain — and, more usefully, it lists the deficiencies examiners keep finding.

That list is the part worth reading twice. DFS describes incomplete asset scope and visibility, weak methodologies, failures to account for evolving risk, insufficient governance and risk treatment, and cybersecurity programs that do not actually reflect the risks the assessment identified.

Read alongside the substance of the guidance, four demands stand out. Each one is a question a regulator can ask in an examination. Most firms cannot answer any of them from the documents they currently maintain.

1. Where does the sensitive data live, and who reaches it?

The guidance is direct about what feeds the assessment. Covered entities should consider, as it puts it, “where NPI resides, how it flows through Information Systems, who has access to assets, and how assets are protected.”

That is four distinct facts about every asset, and they have to be true at the same time. Most firms can produce one of them on demand — usually a list of machines — and reconstruct the rest by interview. An inventory that names the asset but not who logs into it, or names the asset but not which controls are actually running on it, does not support the sentence DFS is asking you to write.

This is where Section 500.13’s asset-inventory requirement stops being a standalone checkbox and becomes load-bearing. We wrote about that requirement on its own here. The September guidance makes the consequence explicit: the inventory is an input to the risk assessment, and a risk assessment built on an incomplete inventory inherits every gap in it.

2. Which third parties, ranked how?

The guidance asks entities to assess third-party risk against four specific dimensions: the criticality of the services provided, the sensitivity of the information accessed or maintained, the level of connectivity to your information systems, and the potential operational impact.

Notice what is absent from that list. There is no mention of the vendor’s security-questionnaire score, their SOC 2, or their published security page. Every dimension DFS names is a fact about your environment — what the vendor touches, how deeply it is wired in, and what breaks if it fails.

A questionnaire cannot produce any of them. It describes the vendor. DFS is asking about the relationship.

3. Where are the single points of failure?

This is the most interesting demand in the letter, and the one with the least existing tooling behind it. Entities are asked to assess cyber interdependencies among information systems and shared technology platforms, and to identify potential single points of failure where a concentration of services creates systemic risk.

The guidance makes the reasoning plain: technologies, platforms, or vendors that present limited risk when evaluated independently may collectively create significant risk when multiple critical systems or business functions rely on common infrastructure.

That sentence describes something no vendor-by-vendor review will ever surface. Assess ten suppliers one at a time and each comes back acceptable. The risk is in the overlap — and the overlap is invisible unless you can see which of your systems depend on the same thing underneath.

This has become sharper, not easier, as AI has spread through the software estate. A firm may have approved a dozen SaaS applications over several years, each on its own merits. What none of those reviews asked is which model providers those vendors route customer data to, because in most cases the vendors had not yet disclosed any. Several of those applications may now share a small number of providers underneath. Nothing in the procurement record shows it.

4. What is AI doing to your risk?

The guidance names “the adoption or use of artificial intelligence” among the emerging risks that should prompt a reassessment, and points to DFS’s separate guidance on risks associated with frontier AI models as a trigger for updating an assessment.

The practical difficulty is that “adoption” is no longer a decision with a date. AI arrives three ways, and only one of them goes through a procurement process: tools staff sign up for directly, AI features switched on inside applications you already approved, and model runtimes installed locally on endpoints. The first is what most shadow-AI tooling looks for. The second and third leave no signup, no grant, and no new vendor record — so a risk assessment that asks “which AI did we adopt” will confidently return an answer that is mostly wrong.

The thread running through all four

None of these are policy problems. They are the same problem wearing four costumes: the firm cannot see its own estate accurately enough to reason about it.

DFS asks for “a clear and repeatable methodology,” for “reasonable estimates of the likelihood and impact of risks,” and for documentation showing how risks were identified, assessed and addressed — including the methodology used, the data considered, and the rationale behind the conclusions. That standard is achievable. But it is only achievable on top of facts. A methodology applied to an estimated inventory produces an estimated risk assessment, documented to the same standard as a guess.

Where KeyCaliber fits

KeyCaliber builds the picture these four questions need from the security tools you already run — EDR, vulnerability scanners, CMDB, identity, network and gateway telemetry — connected by API, with nothing to deploy.

  • One reconciled inventory, continuously current. Every asset, assembled from the systems that already hold the facts, with owner, location and classification carried through from the sources that maintain them.
  • Criticality computed, not asserted. KeyCaliber derives what each asset is worth to the business from how it behaves — what runs on it, who signs into it, what it communicates with — rather than reading a tag someone set in a CMDB years ago. That is the “criticality” and “potential operational impact” dimension DFS asks you to rank third parties against.
  • How assets are protected. Coverage is tracked per asset across roughly thirty control categories, with applicability rules so a control that cannot run on an asset is not counted as a gap. That answers the fourth clause of the NPI sentence directly, and it is the same evidence Section 500.14 asks for.
  • Who reaches what. Identity-to-asset relationships are a first-class part of the model, so — where your identity data supports it — the inventory can say who signs into a given system, not only that the system exists.
  • Who your vendors hand data to. For each SaaS and AI vendor in your estate, KeyCaliber reads the vendor’s own published sub-processor list and records which of those third parties are AI model providers, dated. Where a vendor publishes nothing, and where a list exists but cannot be read, are recorded as two different states — because they mean two different things.
  • The AI you didn’t approve, including the kind nobody signs up for. Local model runtimes found through the software inventory your EDR and MDM already collect, AI features inside applications you have already approved, and hosted AI reached over the network — each with the disposition you have recorded against it and a full history of who changed it and when.
  • Evidence you can hand over. Inventories and the AI system register export as dated files whose header states what the export covers and which discovery sources were live when it ran — so a row missing because a feed was down never reads as a row missing because there was nothing there.

DFS did not ask for a new document. It asked for a risk assessment that is traceable to facts about the estate, reviewed when the estate materially changes. The firms that struggle with this guidance will not struggle with the methodology. They will struggle because the facts underneath it were never assembled in one place.

The inventory was always the hard part. This guidance just made it the examinable part.


← All articles