---
title: "CISO Security Checklist for Workday AI Connectors — Kiweely"
description: "Ten concrete InfoSec questions for connecting Claude, ChatGPT, or Gemini to Workday: OAuth grant type, token sealing, payload logging, write cards, and IP…"
url: https://kiweely.com/blog/ciso-security-review-checklist-connecting-llms-to-workday
---

# The CISO Checklist for Connecting Claude, ChatGPT, or Gemini to Workday

Before approving a Workday connector for Claude, ChatGPT, or Gemini, CISOs and security architects should verify ten concrete controls across OAuth grant types, token storage, payload logging, out-of-band write approval, field redaction, IP allowlisting, and Workday audit trails.

-   By Kiweely
-   November 15
-   6 min read

Employees already ask Claude, ChatGPT, and Gemini Workday questions—and when IT blocks a live connector, staff export custom reports to CSV on their laptops and paste sensitive worker rows straight into chat outside Workday security. Approving a governed, user-delegated MCP connector eliminates shadow spreadsheet uploads, provided the connector passes **ten non-negotiable architectural checks** before touching your Production tenant.

Use the ten questions below as a pass/fail gate during security architecture review—whether you are evaluating [Workday Agent Gateway, a multi-app broker, an internal engineering build, or Kiweely Bridge](https://kiweely.com/bridge/compare).

## Quick-Reference CISO Scorecard

| # | Control Area | Reject Immediately If... | Require Proof Of... |
| --- | --- | --- | --- |
|  **1** |  **OAuth Grant Type** | Uses a shared Integration System User (`Register API Client for Integrations`) | `Register API Client` with Authorization Code Grant + PKCE per worker |
|  **2** |  **SSO & MFA Path** | Prompts users for credentials or bypasses IdP MFA policies | Direct browser sign-in on Workday enforcing SAML/OIDC SSO & MFA |
|  **3** |  **Credential Storage** | Central DB table storing employee refresh/access tokens at rest | Zero-table sealed client-held grants (AES-256-GCM) decrypted in RAM only |
|  **4** |  **Payload Persistence** | Stores `action_logs`, prompts, or cached worker profiles on disk/DB | Stateless RAM pass-through + schema-locked metadata-only logs |
|  **5** |  **Field-Level Redaction** | Passes raw SSNs, bank numbers, or peer pay to LLMs when role permits | Deterministic in-memory field stripping before tool output leaves gateway |
|  **6** |  **Default Write Posture** | Exposes write/mutation tools on day one or toggles via prompt text | Read-only by default; disabled write tools omitted from `tools/list` |
|  **7** |  **Write Confirmation** | Relies on in-chat LLM *“Are you sure?”* prompts vulnerable to injection | Out-of-band browser approval card binding SHA-256 payload hash (*Now* vs. *After*) |
|  **8** |  **Network Perimeter** | Shared dynamic egress IPs with no tenant IP lock option | Static outbound IP locked in `Maintain IP Ranges` or self-hosted VPC container |
|  **9** |  **Tenant Audit Trail** | Attributes reads/writes to one bot account or keeps audit logs only outside Workday | Native per-worker entries in Workday `View User Activity` & BP audit trails |
|  **10** |  **Desktop Footprint & Upkeep** | Pushes local Python/Node runtimes, `.mcpb` bundles, or laptop secrets via MDM | Zero-install remote HTTPS MCP URL with managed semiannual Workday updates |

## The 10 Questions in Detail

### 1\. Does every API call run under a user-delegated OAuth grant rather than an ISU?

Ask which Workday task registers the connection. If the setup guide tells your admin to run [Register API Client for Integrations and create an Integration System User](https://kiweely.com/guides/workday-integration-system-user-api-client), stop the review. Shared `ISU` credentials bypass Workday’s row- and field-level security model and force third-party code to emulate Workday org hierarchies. Require **Register API Client** with **Authorization Code Grant** and **Support Proof Key for Code Exchange (PKCE)** ([setup walkthrough](https://kiweely.com/guides/workday-api-client-pkce-sso-without-workday-password)) so Workday evaluates every REST, WQL, SOAP, and RaaS request against the signed-in employee’s own security groups.

### 2\. Does employee sign-in enforce our corporate SSO, Conditional Access, and MFA?

The connector should never see, proxy, or touch an employee’s password or MFA factor. During the OAuth 2.0 Authorization Code flow, the employee’s browser authenticates directly on Workday’s sign-in endpoint—meaning your existing Okta, Microsoft Entra ID, or Ping SAML/OIDC policies, device posture checks, and Workday MFA rules execute unchanged.

### 3\. Where do employee Workday OAuth tokens live when no query is running?

Even user-delegated SaaS brokers create a massive blast radius if they hoard thousands of active HR and executive tokens in a central PostgreSQL or Redis table. Verify whether the connector supports **sealed client-held tokens**—encrypting each user’s grant with AES-256-GCM inside the opaque session token held by the MCP client, so idle gateway servers and databases hold zero customer credentials.

### 4\. Can you prove zero-payload logging with your structured log schema and DDL?

Do not accept *“We don’t train on your data”* as an answer about gateway storage. Inspect the structured telemetry schema emitted on `stdout`. A zero-payload connector logs only scalar operational fields (`timestamp`, `tenant_id`, `tool_name`, `http_status`, `latency_ms`) and contains zero database columns for WQL strings, RaaS rows, prompts, or responses.

### 5\. How do you stop HR Partners from accidentally pulling SSNs or bank routing numbers into chat?

Per-user OAuth respects Workday domains, which means an `HR Partner` or `Payroll Administrator` who legitimately holds `Person Data: ID Information` or `Worker Data: Payroll (Payment Elections)` in Workday can inadvertently pull Social Security Numbers or bank account numbers into an LLM context window when running a broad roster query. Require deterministic **in-memory field stripping** that removes blocked fields in RAM before the assistant ever receives the tool output.

### 6\. Is the connector strictly read-only by default, with disabled tools hidden from the model?

Write tools that are “disabled” by returning an error after the LLM tries to call them waste context and confuse employees. When write tasks are off, they must be completely absent from the MCP `tools/list` schema presented to Claude, ChatGPT, and Gemini—and enabled one task at a time by the tenant admin.

### 7\. Are enabled writes gated by an out-of-band browser approval card?

Indirect prompt injection inside an email, resume, or PDF can instruct an LLM to call a write tool and pass `confirmed: true` without human consent. Any enabled Workday write (such as `Request Time Off` or `Change My Business Title`) must pause on the server and require the authenticated employee to click **Allow** on an out-of-band browser card showing human-readable **Now** vs. **After** values locked to a pending action hash.

### 8\. Can we lock Workday API traffic to a dedicated static IP or run the container in our own VPC?

For enterprise tenants, verify two network perimeter options:

-   **Static IP allowlisting:** A dedicated outbound IPv4 address you add in Workday’s **Maintain IP Ranges** task and enforce via **Manage Authentication Policies**, causing Workday to reject token replay from any other network.
-   **Self-hosted stateless container:** For zero external network hops, the ability to deploy the signed, stateless container (`readOnlyRootFilesystem: true`, zero volumes) inside your own AWS ECS, GCP Cloud Run, Azure Container Apps, or Kubernetes VPC.

### 9\. Does Workday remain the authoritative audit trail for every read and write?

Verify in Workday Sandbox that **Enable User Activity Logging** is active under **Edit Tenant Setup - System**, then run a test query through the connector and inspect Workday’s **View User Activity** report (and `/activityLogging` REST API). Every read and business process event should appear under the individual employee’s `System Account`—giving your SIEM team full attribution inside Workday’s native audit pipeline. Also [audit over-permissive Workday security groups](https://kiweely.com/guides/workday-security-audit-before-enabling-ai-agents) before go-live (and compare [Workday ASOR security requirements](https://kiweely.com/guides/workday-agent-system-of-record-security) if evaluating Workday-native agents).

### 10\. Does rollout require laptop runtimes, MDM packages, or manual release patching?

Local desktop MCP servers require pushing Node.js or Python runtimes, distributing `.mcpb` bundles via Jamf or Intune, storing credentials on employee disks, and pushing emergency updates twice a year when Workday updates WSDL and REST schemas. Require a **zero-install remote HTTPS MCP server** where workspace admins register one connector URL for Claude, ChatGPT, and Gemini across desktop, browser, and mobile.

## How Kiweely Bridge Answers All 10 Checks

[Kiweely Bridge](https://kiweely.com/bridge) (coming next) was architected specifically around every row of this CISO scorecard ([review the full component-by-component trust matrix at `/bridge/security`](https://kiweely.com/bridge/security)):

-   **User-delegated OAuth + PKCE (`Register API Client`), never an ISU**, with an AI admin agent that configures scopes, sets RAM field-redaction rules, and runs persona boundary tests in Sandbox before Production rollout.
-   **AES-256-GCM sealed client-held tokens, zero stored payloads, schema-locked logs, read-only by default, and out-of-band browser write cards** showing *Now* vs. *After*.
-   **Flexible hosting and transparent pricing:** Deploy on **Kiweely Cloud** ($499/mo up to 1,000 Workday employees, $1,499/mo for 1,001–5,000, $3,999/mo for 5,001–20,000, unlimited calls across Production and all Sandbox/implementation tenants; free for one person on a personal Workday developer tenant), or choose **Dedicated single-tenant** (static outbound IP) or **Self-hosted stateless container** inside your AWS, GCP, Azure, or Kubernetes VPC (Enterprise by quote).

## Frequently asked questions

### Why is pasting exported Workday spreadsheets into Claude or ChatGPT worse than a governed live connector?

Without a live connector, employees export CSVs onto unmanaged laptop folders and upload full tables into chat—stripping Workday domain constraints, leaving stale PII on disk, and producing zero Workday user-activity audit records.

### Should we approve a Workday AI connector that authenticates via an Integration System User (ISU)?

No. An ISU shares one background service identity across all employees, forcing external code to guess Workday row and field permissions and attributing every transaction in Workday logs to the bot account rather than the human worker.

### Does connecting Claude, ChatGPT, or Gemini safely require buying Workday's own AI SKUs?

No. Any standard Workday tenant supports user-delegated OAuth 2.0 via Register API Client (Authorization Code Grant with PKCE), Maintain IP Ranges, Manage Authentication Policies, and View User Activity without Flex Credits or an Extend license.

### What is the fastest deal-breaker test in a CISO architecture review?

Ask for the vendor's database DDL for OAuth tokens and action history plus a raw sample structured log line. If the vendor stores plaintext/column-encrypted refresh tokens for all workers in a central table or logs tool output JSON, treat the gateway as a secondary HR system of record.

## Read next

[All posts](https://kiweely.com/blog)

-   [November 15 ### Using Face ID and Fingerprint Unlock with Workday Single Sign-On for Deskless Workers](https://kiweely.com/blog/biometric-face-id-fingerprint-reauth-for-workday-mobile-apps)
-   [November 15 ### Pass-Through Workday Screens vs. Caching HR Portals: Security and Sync Tradeoffs](https://kiweely.com/blog/pass-through-workday-ui-vs-caching-hr-portals)
-   [November 15 ### Five Things to Check in Your Workday Sandbox Before Connecting External AI or Custom Screens](https://kiweely.com/blog/preparing-your-workday-sandbox-for-kiweely-bridge-and-leaf)

### More guides on Business processes and security

[All 27 guides](https://kiweely.com/guides#business-processes)

-   [What does rescinded mean in Workday? Rescind vs cancel vs correct, and when each applies](https://kiweely.com/guides/workday-business-process-rescind-cancel-correct)
-   [Workday custom validations and condition rules: block bad data with Critical or Warning severity](https://kiweely.com/guides/workday-custom-validations-condition-rules)
-   [Workday Role-Based, User-Based, Intersection, and Aggregation Security Groups](https://kiweely.com/guides/workday-role-based-user-based-intersection-security-groups)
-   [Workday Domain Security vs. Business Process Security Policies: Which Controls What](https://kiweely.com/guides/workday-domain-vs-business-process-security-policies)
-   [Workday Business Process Advanced Routing: Exclude Initiator, Prior Approvers, and Alternates](https://kiweely.com/guides/workday-business-process-routing-mutual-exclusion)
-   [Constrained vs unconstrained security groups in Workday: which workers can a member see?](https://kiweely.com/guides/workday-constrained-vs-unconstrained-security-groups)
-   [Workday Awaiting Action: who is the step waiting on, and how do you move it?](https://kiweely.com/guides/business-process-awaiting-action)
-   [My Workday business process condition rule is not working and the step did not fire: how to troubleshoot it](https://kiweely.com/guides/condition-rule-not-working)

## Your Workday work, done by AI.

Kiweely Vine is the AI agent that builds, deploys and tests Extend apps as real people, makes changes in any Workday task screen, answers from your real tenant and builds reports.

[Start for free](https://kiweely.com/download) [See how it works](https://kiweely.com/vine/how-it-works)

## Get the next one in your inbox.

One email a month for Workday teams: what’s changing in Workday and what to retest, one tip tried in a real tenant, and what’s new at Kiweely.
