New

AI Connector & Plugin Security Review Checklist

Get Claude connectors and ChatGPT plugins through security, data protection and architecture review. A structured checklist for both scenarios: connectors into approved systems, and plugins with their own backends.

What's inside this template

Who it's for

AI Operations Leads, CIOs, information security teams, data protection officers, system owners and enterprise architects reviewing Claude connectors, ChatGPT plugins or other AI tool integrations

When to use

Before submitting any Claude connector or ChatGPT plugin for approval; when building a review pack for your security team; when a connector is stuck in review and you need to know what's missing

Key benefit

A two-part checklist that separates connector review (scopes, roles, authentication, logging) from vendor review (for plugins with their own backends), so you submit a complete pack rather than answering questions one at a time

Sections included

  • Scenario identification: which review applies
  • Part A: Connector review (ownership, implementer questions, AI vendor questions, data protection, logging, pilot)
  • Part B: Vendor and backend review (data and hosting, assurance, contract and supplier onboarding)
  • Implementation guidance

Complete template content

Prefer Google Docs? Open this template in Google Docs to make a copy and customise it directly. Open in Google Docs →

NOTE: To use this checklist, copy the content using the “Copy page” button above, then fill in each line before you submit for security review. Read the companion guide on connector security reviews for the full context behind each section.


AI connector & plugin security review checklist

A free template for getting a Claude connector or ChatGPT plugin approved. Fill in each line before you submit for security review.

0. Which scenario?

  • Scenario 1: connector into a system you’ve already approved (e.g. NetSuite, HubSpot, Xero). Complete Part A.
  • Scenario 2: plugin with its own backend from a vendor you haven’t approved. Complete Part A and Part B.

Part A: Connector review (every connector)

A1. Ownership

  • Decision owner named:
  • Reviewers: information security, DPO, system owner, enterprise architecture (+ procurement/third-party risk for scenario 2)
  • Business owner and one-sentence use case:

A2. Questions for your implementer

  • Scopes required, each justified; unused scopes removed
  • Records read
  • Records written (if any)
  • Roles that get the connector
  • Connected account’s permissions inside the system of record
  • Authentication model: individual OAuth / managed via identity provider
  • Joiner, leaver and revocation process
  • Post-go-live support owner

A3. Questions for the AI platform vendor

  • Processing regions
  • Training on customer data (yes/no, evidence link)
  • Default retention and configurability
  • Audit events and compliance exports available
  • DPA and trust centre links

A4. Data protection

  • Data flow diagram, personal and special category data marked
  • Lawful basis
  • DPIA screening done; full DPIA if required
  • Subject access request handling
  • Retention on AI platform aligned to policy

A5. Logging and audit

  • AI platform audit logs: where, who reviews, how often
  • Source system logs
  • Actions traceable to a named user

A6. Pilot

  • Read-only scopes only
  • Pilot group limited by role
  • Duration:
  • Exit criteria (usage, accuracy, incidents, log review)
  • Write-back submitted as a separate change with pilot evidence

Part B: Vendor and backend review (scenario 2 only)

B1. Data and hosting

  • What data the plugin’s backend stores
  • Where it is hosted (region)
  • Retention period and deletion on request
  • Separation from other customers’ data
  • Sub-processors list

B2. Assurance

  • Current security evidence (and its date)
  • Model training on your data (vendor’s or any third party’s)
  • Incident notification and support contacts

B3. Contract and supplier onboarding

  • Data processing agreement
  • International transfer mechanism (if data leaves the UK)
  • DPIA covers the vendor’s processing
  • Exit and data return/deletion terms
  • Added to supplier register

Implementation notes

Before you start

Name the decision owner first. Without one named owner (typically a CISO’s deputy or enterprise architecture lead), each reviewer waits for the others. Get their agreement to own the decision before you write anything else.

Know which scenario you’re in. If the connector reaches into a system your organisation already contracted and approved (NetSuite, HubSpot, Xero, your own M365 tenant), that’s scenario 1. If the plugin brings its own backend on its own hosting with its own sub-processors, that’s scenario 2. Scenario-2 plugins are arriving through ChatGPT’s and Claude’s plugin directories, and they need a full vendor review.

Drawing the data flow diagram

Use arrows to show user → AI platform → connector → source system. Mark which flows carry personal data and which carry special category data. This diagram is the artefact your DPO needs for DPIA screening, and security teams approve diagrams faster than prose.

Justifying scopes

List every permission the connector requests. Explain which part of the workflow needs each one. Remove anything unused. Both Claude and ChatGPT let admins restrict scopes centrally now, so “it asks for everything” is not acceptable. Also check the connected account’s permissions inside the system of record, because the connector acts with those permissions.

Piloting read-only

Start every connector read-only. Reading a HubSpot deal is a different risk from updating it. Run the pilot with one team (using role-based access to limit who gets the connector), set a duration (typically four weeks), and define exit criteria: usage threshold, accuracy checks, no incidents, and a clean log review. Then submit write actions as a separate, smaller review with the phase-one evidence attached.

Vendor review for scenario 2

Treat the plugin provider like any new SaaS supplier: data processing agreement, international transfer mechanism if data leaves the UK, a DPIA that covers the vendor’s processing, exit and data deletion terms, and a place on your supplier register. Don’t let the plugin directory’s install button skip procurement steps that a normal SaaS contract would never skip.

When to split the reviews

Don’t batch a scenario-1 connector with a scenario-2 plugin. Scenario 2’s vendor checks will hold up scenario 1’s faster approval. Split them and move the simpler one first.


Template by Kowalah, UK Anthropic Implementation Specialist. kowalah.com

Get started

How to use this checklist

01

Identify your scenario

Scenario 1: connector into an already-approved system (e.g. NetSuite, HubSpot). Complete Part A only. Scenario 2: plugin with its own backend from an unapproved vendor. Complete Part A and Part B.

02

Name the reviewers up front

Before filling anything in, name the decision owner and every reviewer (information security, DPO, system owner, enterprise architecture; add procurement for scenario 2). One named owner stops circular waits.

03

Fill in Part A for every connector

Document scopes (each justified, unused ones removed), roles that get the connector, authentication model, logging, DPIA screening, and a read-only pilot with exit criteria. Draw a data flow diagram with personal data marked.

04

Add Part B for scenario 2 (new vendor)

For plugins with their own backend, add vendor checks: what data the backend stores and where, retention, sub-processors, current security evidence, DPA, transfer mechanism, and supplier register entry.

05

Pilot read-only before write-back

Start every connector read-only. Run the pilot with role-based access to one team, set exit criteria (usage, accuracy, no incidents, clean log review), then submit write actions as a separate change with phase-one evidence.

Questions

Frequently asked questions

Common questions about connector security reviews

What's the difference between scenario 1 and scenario 2?
Scenario 1 is a connector into a system you've already approved and contracted (e.g. Claude reaching into your NetSuite instance). You're approving the connector's scopes and roles, not the system. Scenario 2 is a plugin that brings its own backend, storing data on its own hosting with its own sub-processors. Approving it means approving a new vendor and a new application, not just a connector. ChatGPT's plugin directory and Claude's plugin directory both include scenario-2 plugins, and they require a full vendor review.
Why do I need a named decision owner?
Without one named owner, each reviewer waits for the others. A connector stuck for months is usually stuck because no one has final sign-off authority. Name the decision owner (typically a CISO's deputy or enterprise architecture lead) before you write the pack, get their agreement, and direct every question to them.
What should the data flow diagram show?
User, AI platform (Claude or ChatGPT), connector, and source system, with arrows for what moves where. Mark personal data and special category data on each flow. Security teams approve diagrams faster than prose, and the diagram is the artefact the DPO needs for DPIA screening.
How do I justify scopes?
List every permission the connector requests and explain which part of the workflow needs it. Remove anything unused. Both Claude and ChatGPT now let admins restrict scopes centrally, so 'it asks for everything' is no longer acceptable. Check the connected account's permissions inside the system of record too, because the connector acts with those permissions.
What's a read-only pilot?
A connector with read scopes only (no write, update or delete), limited to one team via role-based access, with a fixed duration (typically four weeks) and exit criteria (usage threshold, accuracy checks, no incidents, clean log review). Write-back (e.g. updating a HubSpot deal or posting a Slack message) goes to review separately after phase one, with the pilot evidence attached.
What evidence do I need from the AI vendor?
Processing regions, whether customer data is used for model training, default retention periods and whether you can change them, what audit events and compliance exports exist, and links to the DPA and trust centre. Most of this is public (Anthropic's trust centre, OpenAI's trust centre, their published DPAs). Collect the links rather than rewriting policy that the vendor maintains.
When does a connector need a DPIA?
When it processes personal data at scale, in new ways, or involves special category data. The threshold is set by UK GDPR Article 35. Start with DPIA screening (a short yes/no checklist, often published by your own DPO) and escalate to a full DPIA if screening says so. Bring the DPO in when you draw the data flow diagram, not after the connector is built.
What counts as a scenario-2 vendor review?
Treat the plugin provider like any new SaaS supplier: data processing agreement, international transfer mechanism if data leaves the UK, a DPIA that covers the vendor's processing, exit and data deletion terms, and a place on your supplier register. Don't let the plugin directory's install button skip procurement steps that a normal SaaS contract would never skip.
Can I batch multiple connectors into one review?
Only if they're all scenario 1 (same approved system, similar scopes, same roles). Don't batch a scenario-1 connector with a scenario-2 plugin, because the vendor checks for scenario 2 will hold up the faster scenario-1 review. Split them and move the simpler one first.

Ready when you are

We've been through this review ourselves

Kowalah is a UK Anthropic Implementation Specialist, and our own plugin and platform go through exactly this security review. If you'd like our security answers to start your pack, or help building the review for your own connectors, book a conversation.

Book a Conversation