Cyber Essentials v3.3 never mentions AI tools. An AI service such as ChatGPT or Copilot Chat that staff sign in to with a work account and use with organisational data fits the v3.3 definition of a cloud service in our reading, so it is in scope: list it, apply MFA for users and administrators, and control who has accounts.
Key takeaways
- The NCSC Requirements for IT Infrastructure v3.3 do not name AI tools; this guide applies the v3.3 cloud services rules to them, and that application is our interpretation.
- v3.3 defines a cloud service as on-demand, scalable, hosted on shared infrastructure, accessed via an account and storing or processing your organisation's data, and says cloud services cannot be excluded from scope.
- For SaaS, the v3.3 responsibility table puts user access control on your organisation, so MFA for every user and administrator (A7.16, A7.17) is your job, and a no to either fails automatically.
- AI browser extensions and desktop apps are software under v3.3, which lists extensions in its definition, so they must be licensed, supported and updated within 14 days for critical fixes (A6.5).
- Named in the requirements?
- No. AI tools are not mentioned in v3.3 or the Danzell question set
- How they are treated (our interpretation)
- As SaaS cloud services when staff use them with work accounts and organisational data
- Must list
- A2.9, every cloud service; cloud services cannot be excluded from scope
- Automatic fails
- A7.16 and A7.17 (MFA for administrators and users); A7.15 if you wrongly list a service as having no MFA
- Extensions and apps
- Software under v3.3: licensed, supported, updated within 14 days (A6.2, A6.5)
Staff are using AI assistants, whether or not there is a policy for it. The question we get during assessments is whether those tools belong in the Cyber Essentials scope. The short answer is that the requirements never mention AI, but they do define a cloud service, and most AI tools used for work fit that definition. This guide applies the v3.3 rules, labels clearly where we are interpreting, and lists the questions that are affected. Facts were checked on 9 November 2026.
Do the Cyber Essentials requirements mention AI tools?
No: neither the NCSC Requirements for IT Infrastructure v3.3 nor the Danzell question set names ChatGPT, Copilot, Gemini or AI in general.
That is not a loophole. The scheme is written in terms of devices, software, accounts and cloud services, and leaves you to decide which category each tool falls into. Every AI tool your staff use falls into at least one of them.
| How the tool is used | v3.3 category | Main questions |
|---|---|---|
| Web or mobile AI chat signed in with a work account | Cloud service (SaaS) | A2.9, A7.14 to A7.17, A7.1, A5.2 |
| AI features inside an existing SaaS, such as Copilot in Microsoft 365 | Part of that cloud service | Covered by that service's MFA and accounts |
| AI browser extension | Software (v3.3 lists "extensions") | A6.2, A6.5, A8.5 |
| AI desktop app or coding assistant installed on a laptop | Software | A6.2, A6.5, A5.1 |
| AI app or meeting note-taker granted access to your Microsoft 365 or Google data (OAuth consent) | A separate cloud service, plus a third-party app permission in your tenant | A2.9, A7.1, A5.1; restrict which apps users can consent to |
| AI model provider's API platform your developers build on | Usually SaaS or PaaS cloud service; IaaS only if you run the model on your own virtual machines | A2.9, A7.16, plus the software development section |
The table above is how we read the requirements; neither v3.3 nor the Danzell question set contains AI-specific guidance. If your Certification Body reads a case differently, its view decides your assessment, so raise unusual AI tools with your assessor before you submit.
Is an AI chat service a cloud service under Cyber Essentials?
Yes, in our reading: when staff sign in with a work account and use it with organisational data, it meets the v3.3 definition of a cloud service.
The scheme requires you to apply this definition from v3.3: a cloud service is "an on-demand, scalable service, hosted on shared infrastructure, and accessible via the internet". For Cyber Essentials, it "will be accessed via an account (which may be credentials issued by your organisation, or an email address used for business purposes), and will store or process data for your organisation." Question A7.14 prints the same definition.
A ChatGPT, Claude, Gemini or Copilot Chat account opened with a work email and used to draft documents or summarise files meets each part. v3.3 then states that if your data or services are hosted on cloud services, "these services must be in scope. Cloud services cannot be excluded from scope." Question A2.9 asks you to list them all, and the question set adds that social media accounts count too, so the bar for what counts is low.
Who is responsible for which controls on an AI service?
For SaaS, v3.3 says user access control is typically yours, configuration is shared, and firewalls, updates and malware protection are typically the provider's.
| Control | Who typically implements it for SaaS (v3.3 Table 1) | What it means for an AI tool |
|---|---|---|
| Firewalls | The cloud provider | Usually nothing for you to configure |
| Secure configuration | Both your organisation and the cloud provider | Turn off features and connectors you do not use (A5.1) |
| Security update management | The cloud provider | Usually nothing for you to patch in the service itself |
| User access control | Your organisation | Approved accounts, least privilege, MFA, leavers removed |
| Malware protection | The cloud provider | Usually nothing for you to install in the service |
v3.3 also says "the applicant organisation is always responsible for ensuring all controls are implemented". Where the provider implements a control for you, v3.3 says you must make sure the provider "has committed to implementing this via contractual clauses or documents referenced by contract", such as published security statements.
What do you have to do for MFA on AI tools?
The scheme requires MFA for every user and every administrator of every cloud service that offers it, and a no to either question is an automatic fail.
A7.16AUTO-FAIL covers administrators and A7.17AUTO-FAIL covers users. The question set says MFA must be used "in conjunction with a password of at least 8 characters", and that "the assessment will fail if the answer is no." If an AI tool genuinely offers no MFA, list it in A7.15AUTO-FAIL, but that question also fails automatically if a service you list does offer MFA.
In practice there are two routes. One is single sign-on through your identity provider, so your existing MFA policy applies to the AI tool. The other is the tool's own MFA setting, which each user turns on. The table shows what each vendor's help pages said on 9 November 2026; plans and settings change, so recheck before you submit.
| Service | MFA for users | Can an administrator enforce MFA? | Single sign-on | Vendor source |
|---|---|---|---|---|
| ChatGPT (all plans) | Yes: authenticator app, push, text or WhatsApp code, or passkey; OpenAI says options "may vary depending on your device, country, account tier" | No: "MFA cannot currently be enforced at the ChatGPT workspace or API Platform organization level" | ChatGPT Business includes SSO (SAML or OIDC), and a workspace can require it for verified domains; Enterprise configures SSO in Admin Console | MFA, Business SSO |
| Microsoft Copilot Chat and Microsoft 365 Copilot | Through the user's Microsoft Entra account | Yes: a Conditional Access policy can require MFA for all users and all resources | Microsoft tells admins to have users "sign in with their Microsoft Entra account" | Copilot, Conditional Access |
| Gemini app with Google Workspace | Google 2-Step Verification on the Workspace account | Yes: Workspace admins can turn on 2-Step Verification enforcement | Workspace sign-in; admins control access to gemini.google.com, the Gemini mobile apps and Gemini in Chrome | 2-Step Verification, Gemini app |
| Claude Team and Enterprise | Through your identity provider once SSO is set up | Yes, through your identity provider, once SSO is "explicitly enforced" | "Single sign-on is available for Team plans, Enterprise plans, and Console organizations" | Claude SSO |
Because OpenAI does not let administrators enforce ChatGPT MFA, a business ChatGPT workspace is easiest to evidence with SSO required for your domain, so your identity provider's MFA applies. Without SSO, you need each user to switch MFA on and a way to show it is on for everyone, for example screenshots from each account's security settings.
A team buys a few AI subscriptions on a company card, each with its own login and no MFA. Nobody lists them in A2.9 because IT did not set them up. An assessor who asks about AI use finds them, and A7.17 is now a no. Find these before you submit: check expense claims and card statements for AI subscriptions, and ask each team what they use.
Which user access control rules apply to AI accounts?
All of them: accounts must be approved, unique to a person, limited to what the job needs, and removed when no longer required.
The scheme requires a process to create and approve user accounts (A7.1), least privilege (A7.4), separate administrator accounts for administration (A7.6), and removal of unnecessary accounts (A5.2). For AI tools that means no shared team login, no admin role for everyday users, and an owner who removes leavers.
Microsoft notes you can control whether users sign in to Copilot with personal Microsoft accounts by using tenant restrictions, rather than blocking the Copilot address. Steering staff to one approved, centrally managed AI service makes the A2.9 list, MFA and leaver process far easier to evidence. The scheme does not require you to block other tools.
Are AI browser extensions and desktop apps in scope?
Yes: v3.3 defines software to include "extensions", so AI extensions and installed apps on in-scope devices must meet the security update management rules.
The scheme requires all software on in-scope devices to be "licensed and supported" and critical or high-risk fixes to be installed within 14 days of release. A6.2 adds that unsupported software must be removed, and "this includes frameworks and extensions". A6.5AUTO-FAIL applies the 14-day rule to "software (including any associated files and extensions)" and is an automatic fail on a no. If you protect devices by limiting installation to approved applications (option B in A8.1), A8.5 requires a current list of approved applications.
An AI extension that reads every page you visit is still just an extension for Cyber Essentials purposes. The questions to answer are whether it is supported by a vendor that issues security fixes, whether it updates automatically, and whether you approved it. Unused ones should go under A5.1, which requires removing software that is not needed.
What about AI services your developers build on?
Platforms your developers use to build AI features are cloud services, usually SaaS or PaaS, so they go in the A2.9 list and their administrator accounts need MFA under A7.16.
v3.3's software development section says "Bespoke and custom components of web applications are out of scope" and points to the UK Government's Software Security Code of Practice; its change log reads "Software Security Code of Practice introduced in Software Development section". Our software development scope page explains where the boundary sits. Cyber Essentials does not assess the AI features you build or the data you send to a model provider; for that, see Cyber Essentials and UK GDPR.
What about AI apps connected to Microsoft 365 or Google?
An AI meeting note-taker, email assistant or agent that a user connects to your Microsoft 365 or Google tenant is a separate cloud service in our reading, and it also holds a permission inside your tenant.
The service itself goes in the A2.9 list and needs MFA on its own accounts (A7.16, A7.17) if it has a login. The permission is the part teams miss: when a user clicks "accept" on a consent screen, the app can read mail, files or calendars without any further sign-in. A7.1 requires a process to approve accounts, and A5.1 asks you to check cloud services and "disable any services that are not required for day-to-day use".
Review the third-party apps already consented to in your tenant, remove the ones nobody approved, and restrict user consent so new AI apps need an administrator's approval. It turns A7.1 and A5.1 into something you can show the assessor.
What should you do before your next assessment?
Find every AI tool in use, decide which are approved, and treat each approved one like any other SaaS.
- Inventory AI services from expense records, single sign-on logs and a short staff survey.
- Add each one used with work data to your A2.9 cloud services list. See cloud services in scope.
- Turn on MFA for every user and administrator, preferably through single sign-on. Follow the Microsoft 365 and Entra ID setup guide if you use Microsoft.
- Remove unsupported or unapproved AI extensions and apps, and turn on automatic updates for the rest.
- Name an owner for each service who approves accounts and removes leavers.
Then check your readiness against the full 2026 question set. If you would rather an assessor reviewed your cloud services list before you submit, Precursor Security's gap analysis starts from £500, and Cyber Essentials from £1,500 with the IASME fee included.
- NCSCCyber Essentials: Requirements for IT Infrastructure v3.3 · Definitions; Scope (iv) Cloud services; User Access ControlRead on ncsc.gov.uk (opens in a new tab)
- Microsoft LearnWhat is Microsoft Copilot?Read on learn.microsoft.com (opens in a new tab)
- help.openai.comOpenAI Help Center, Managing multi-factor authentication (MFA) · Checked 9 November 2026Read on help.openai.com (opens in a new tab)
- help.openai.comOpenAI Help Center, Setting up single sign-on (SSO) for ChatGPT Business · Checked 9 November 2026Read on help.openai.com (opens in a new tab)
- support.claude.comClaude Help Center, Set up single sign-on (SSO) · Checked 9 November 2026Read on support.claude.com (opens in a new tab)
- knowledge.workspace.google.comGoogle Workspace Admin Help, Deploy 2-Step Verification · Checked 9 November 2026Read on knowledge.workspace.google.com (opens in a new tab)
- knowledge.workspace.google.comGoogle Workspace Admin Help, Turn the Gemini app on or off · Checked 9 November 2026Read on knowledge.workspace.google.com (opens in a new tab)
- Microsoft LearnRequire MFA for all users with Conditional Access · Checked 9 November 2026Read on learn.microsoft.com (opens in a new tab)
- IASMECyber Essentials Danzell Question Set 2026 (v16.3)Read on iasme.co.uk (opens in a new tab)
Frequently asked questions
If staff use it for work with an organisational account or business email, yes, in our interpretation. A2.9 asks for all cloud services provided by a third party and says they cannot be excluded from scope. v3.3 defines a cloud service as accessed via an account, which may be a business email address, that stores or processes data for your organisation. A work ChatGPT account fits.
Our view is that Copilot features reached through your Microsoft 365 sign-in are part of the Microsoft 365 service you already list, protected by the same Entra ID account and MFA. A separately purchased AI service with its own login is a separate cloud service. Microsoft says Copilot Chat users should sign in with their Microsoft Entra account to get enterprise data protection.
List it in A7.15 as a service that does not provide MFA, but check carefully first. The question set says A7.15 fails automatically if a service you list does offer MFA. If it truly has none, the password-based authentication rules in v3.3 still apply to the accounts, including protection against brute-force guessing (A7.10).
Yes, in our reading, whoever paid. The v3.3 cloud service definition has two parts: the service is accessed via an account, which 'may be credentials issued by your organisation or an email address used for business purposes', and it stores or processes data for your organisation. A free or paid AI account opened with a work email and used for work meets both, so list it in A2.9 and make sure MFA is on (A7.17).
Possibly. The v3.3 cloud service definition covers accounts using 'credentials issued by your organisation or an email address used for business purposes'. In our reading, a personal address that staff use for work can meet the second part, so the account may be in scope. The safer route is to move that work to an approved work account, list the service in A2.9 and turn on MFA.
It can be. v3.3 defines software to include extensions, and A6.2 says unsupported software, including extensions, must be removed. A6.5 applies the 14-day rule for critical updates to extensions. If you protect devices by limiting installation to approved applications (option B in A8.1), A8.5 requires an approved list. An unmaintained extension from an unknown developer fails all three tests.
No. Cyber Essentials covers technical controls on accounts, devices and software, not what data people share with a service or how the provider uses it. That is a data protection and acceptable use question. See our guide to Cyber Essentials and UK GDPR for what the scheme does and does not cover.
Would you pass the 2026 question set?
Answer 24 yes or no questions, built by Precursor Security, an IASME Cyber Essentials Certification Body, and get a readiness score in about five minutes, with any auto-fail questions you miss flagged first in your results.
- IASME Certification Body
- CREST accredited
- UK offices in Leeds, Newcastle, London and Edinburgh
Help from Precursor Security
Not sure what is in scope?
Scope is set on a free 20-minute call with an assessor before you are quoted. Precursor is an IASME Certification Body and issues the certificate directly.
Share or cite this page