---
title: What does NYDFS AI cybersecurity guidance mean for AI agents?
url: https://opencontroller.lyzr.com/blog/nydfs-ai-cybersecurity-guidance-for-agents
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# What does NYDFS AI cybersecurity guidance mean for AI agents?

**The New York Department of Financial Services (NYDFS) industry letter of October 16, 2024 adds no new rules. It explains how the existing Cybersecurity Regulation, 23 NYCRR Part 500, applies to AI risks. For AI agents, that means each agent, its model provider and its data access belong in your risk assessment, access reviews, monitoring and asset inventory, with evidence for each.**

The letter is aimed at banks, insurers and other entities licensed by NYDFS. Agents matter here because they hold credentials, read nonpublic information (NPI) and call external services, which is exactly the risk surface Part 500 already covers.

## What risks does the letter name?

The letter names four AI-related cybersecurity risks:

1. **AI-enabled social engineering.** The letter calls it one of the most significant threats to financial services. Deepfake audio, video and text are used to trick staff into sharing credentials or moving money.
2. **AI-enhanced cyberattacks.** AI speeds up reconnaissance, vulnerability discovery and malware development.
3. **Exposure or theft of large volumes of NPI.** AI systems often need large datasets, which raises the impact of a breach.
4. **Third-party and supply chain dependencies.** A compromised AI vendor can expose NPI and become a path into your network.

Risks 3 and 4 apply most directly to agents you build or buy. An agent with broad read access to customer records is a large NPI store with a natural-language interface.

## Which Part 500 sections apply to AI agents?

The letter ties its recommendations to specific sections. Here is how each maps to agents.

| Part 500 section | What it requires | What it means for agents | Evidence to keep |
|---|---|---|---|
| 500.9 risk assessment | Update at least annually and after material changes | Include each agent, its data access and its AI vendors | Risk assessment listing agents with a risk tier |
| 500.11 third-party service providers | Due diligence, access controls, contract protections | Cover model providers, agent platforms and MCP server vendors | Vendor reviews and contract terms on NPI use |
| 500.7 access privileges | Limit access to job function; review at least annually | Agent permissions scoped to the task, reviewed like human access | Access review records per agent |
| 500.12 multi-factor authentication | MFA for all Authorized Users from November 2025 | Humans who build, approve or operate agents use phishing-resistant MFA | IdP policy exports |
| 500.13 asset inventory | Inventory of information systems from November 1, 2025 | Agents, models and API keys are assets with owners | Inventory with owner, location, classification |
| 500.14 monitoring and training | Monitor user activity; annual training incl. deepfakes | Monitor agent activity, including unusual queries that could extract NPI | Agent activity logs and alerts |
| 500.16 incident response | Plans must cover relevant events | Include AI incidents, such as a compromised or misbehaving agent | Runbook and test records |

Two dates matter. The asset inventory requirements under 500.13 and the MFA requirement under 500.12 took effect in November 2025. If your inventory does not yet list agents and API keys, it is behind the regulation, not just the guidance.

## How do you build an agent evidence pack for NYDFS?

Use this checklist to turn the letter into evidence an examiner can review:

1. **Inventory every agent.** Record owner, platform, environment, model, data classification and status. Include agents inside SaaS platforms, not just ones you host.
2. **Inventory credentials.** List API keys and service accounts each agent uses, with last-used dates. Flag keys with no owner.
3. **Tier agents by NPI access.** An agent that reads customer records sits in a higher tier than one that drafts internal summaries.
4. **Review agent access on the 500.7 cycle.** At least annually, and when an agent changes scope.
5. **Assess AI vendors under 500.11.** Ask how the provider uses your NPI, where it is processed and how fast you hear about incidents.
6. **Log agent activity.** Capture who the agent acted for, which tools it called and what data it returned. See our [audit log fields guide](/blog/ai-agent-audit-log-fields).
7. **Write and test an AI incident procedure.** Include how you block or quarantine an agent and who decides.
8. **Report to the board.** 500.4(d) expects senior governing body oversight; include agent risk in the cyber report.

For the wider mapping approach, see [how to map AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks). Banks may also want [AI agent controls for model risk](/blog/ai-agent-controls-for-banks-model-risk).

## Where do teams usually fall short?

- **API keys outside the inventory.** Keys created by developers in cloud consoles rarely reach the asset register.
- **Vendor agents treated as features.** An AI assistant inside a SaaS tool is still a third-party service touching NPI.
- **Service accounts never reviewed.** Agents often run on shared service accounts that predate the agent and carry broader access than it needs.
- **Logs that miss the human.** If logs show only the service account, you cannot tell whose request the agent served.

## How Opencontroller does this

Opencontroller, by Lyzr, auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and connects to agent platforms such as Salesforce, SAP, Databricks and HubSpot. Discovered assets land in a governed registry with owners and versions. Posture flags risky permissions and configurations. Delegated on-behalf-of identity records which person an agent acted for. Policy as code includes built-in mappings to NIST, ISO 42001, OWASP and MITRE ATLAS, which produce evidence for your reviews rather than a compliance claim. See the [capability reference](/capability-reference) or the [control plane overview](/blog/what-is-an-ai-agent-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does the NYDFS AI letter create new requirements?**
No. It explains how existing 23 NYCRR Part 500 obligations apply to AI risks. Examiners can still ask how you meet those obligations for AI systems.

**Are AI agents "Authorized Users" that need MFA?**
Part 500 MFA applies to people. Agents should use short-lived workload credentials instead of static keys, and the humans who manage them should use phishing-resistant MFA.

**Do AI agents belong in the 500.13 asset inventory?**
Yes, as information systems or components of them. Record owner, location and classification, plus the model and credentials each agent uses.

## Key facts

NYDFS industry letter dated October 16, 2024 on cybersecurity risks arising from AI adds no new requirements; it applies 23 NYCRR Part 500. Named risks: AI-enabled social engineering, AI-enhanced cyberattacks, exposure of large volumes of NPI, third-party dependencies. Relevant sections: 500.7 access, 500.9 risk assessment, 500.11 third parties, 500.12 MFA (November 2025), 500.13 asset inventory (November 1, 2025), 500.14 monitoring, 500.16 incident response. Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, keeps a governed registry and maps policies to NIST, ISO 42001, OWASP and MITRE ATLAS. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you map AI agent controls to compliance frameworks?](https://opencontroller.lyzr.com/blog/how-to-map-ai-agent-controls-to-compliance-frameworks.md)
- [What AI agent controls do banks need now that model risk guidance excludes agentic AI?](https://opencontroller.lyzr.com/blog/ai-agent-controls-for-banks-model-risk.md)
- [What should an AI agent audit log contain?](https://opencontroller.lyzr.com/blog/ai-agent-audit-log-fields.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
