AI Security Checklist for Small Teams

A small team can manage most AI risk with a short list of habits: know which tools are in use, set clear data rules, keep device access small, control agents and keys, and review it all every quarter. You do not need a security department to start. One owner, one page of rules and a few hours of setup will cover the basics.

This checklist is written for teams of about 5 to 50 people without full-time security staff. Work through it in order. If you handle regulated data, such as health or payment records, get advice for your sector as well, since this list will not cover those rules.

Step 1: Name an owner and list your tools

1. Pick one person to own AI security. This can be a part-time duty.
2. Ask everyone which AI apps, extensions, agents and connected accounts they use. Make it safe to answer honestly, so people do not hide useful tools.
3. Build one shared list with the tool, its owner, its plan type, and what data it touches.
4. Mark each tool as approved, under review or not allowed.
5. Check each provider's terms for data retention and model training. Business and personal plans can differ, so confirm yours.

If you want a framework behind this, the NIST AI Risk Management Framework is intended for voluntary use and has a generative AI profile. A small team does not need to adopt all of it, but it is a useful reference for your owner.

Step 2: Write simple data rules

Keep this to one page. People follow short rules.

• Fine to use: public information and your own drafts with nothing private in them.
• Approved tools only: internal documents and code.
• Never paste: passwords, API keys, customer personal data, health or payment data, and anything under a confidentiality agreement, unless a named tool has been approved for it.
• Always check: a person reviews AI output before it reaches a customer, ships as code or drives a decision about someone.

Add one line on where to ask questions and how to report a mistake without blame.

Step 3: Sign-in and accounts

1. Use company accounts for AI tools, not personal ones.
2. Use single sign-on where your plan offers it.
3. Turn on multi-factor authentication for every AI account.
4. Review connected apps in your email, calendar, drive and code accounts. Remove AI connections nobody uses.
5. Give connections the smallest scope that works, and prefer read-only.

Step 4: Device permissions

Every laptop should have a review when it is set up, and again each quarter.

Mac

1. Open Apple menu > System Settings > Privacy & Security.
2. Review Full Disk Access, Accessibility, Screen & System Audio Recording, Input Monitoring and Automation.
3. Turn off access for AI apps that do not need it.

Our guide to what AI apps can access on your Mac covers each list.

Windows

1. Open Settings > Privacy & security and review the app permission pages.
2. Check Settings > Privacy & security > Recall & snapshots on Copilot+ PCs.
3. Remove unused AI apps from Settings > Apps > Installed apps.

See the full Windows audit guide for detail. On managed devices, Microsoft removes Recall by default, and admins can control it with the Allow Recall to be enabled policy.

Phones

• iPhone: open Settings > Privacy & Security and review each list for AI apps.
• Android: open Settings > Security & privacy > Privacy > Permission manager. Labels vary by phone. Also review Settings > Accessibility, because accessibility access can read your screen.

For a quick first pass across a team, one way to check this is Prycheck. It is a local app for Mac, Windows and Android, with a checklist for iPhone, that shows what AI tools can reach and sends nothing anywhere. It does not show what apps actually do or send, and it does not replace your antivirus or endpoint tools.

Step 5: Browser extensions

1. Choose a short list of approved AI extensions.
2. Ask staff to remove others, or set them to run only on click.
3. Keep sensitive work, such as finance and HR, in a browser profile without AI extensions.
4. If you manage your devices, enforce this with policy. In Edge, blocking all with * and adding an allow list limits installs to approved items. Chrome offers the ExtensionSettings policy for the same job.

Step 6: Agents and MCP servers

OWASP lists excessive agency among its top risks for AI applications. In plain terms, an agent with too much power can do too much damage. It also lists prompt injection first, and no method reliably prevents that.

1. Keep a list of approved MCP servers, with exact versions.
2. Review each server's command, source and token scopes before anyone adds it.
3. Keep tool approvals on. Do not let an agent run commands or send data unattended.
4. Limit agents to the project folder they need.
5. Use read-only credentials by default.
6. Do not point an agent at production systems from a laptop. Use test data and a separate account.
7. Review shared configs, such as .mcp.json, in code review like any other code.

Our MCP security guide has a full review process.

Step 7: API keys and spending

1. Give each person and each project its own key. Avoid shared keys.
2. Keep keys in a password manager or secrets tool, not in chat, documents or code.
3. Turn on secret scanning in your code host, and add a pre-commit scan.
4. Set spending limits with each provider, so a leaked key has a cost ceiling.
5. Rotate keys on a set schedule, and always when someone leaves.

Our API key guide explains how to find and rotate keys.

Step 8: Offboarding and incidents

When someone leaves:

1. Remove their accounts from every AI tool.
2. Revoke their keys and tokens, including any used in shared projects.
3. Remove their connected apps and agent access.
4. Reclaim shared workspaces and saved chats.

Write a short plan for when something goes wrong, such as a key leak or private data pasted in the wrong tool:

1. Tell the owner right away, with no blame.
2. Revoke any exposed credentials.
3. Check the provider's logs and usage.
4. Ask the provider about deleting the data, if that applies.
5. Check whether you have legal or customer notice duties.
6. Write down what happened and change one thing to prevent a repeat.

Step 9: Review every quarter

Put a one-hour meeting on the calendar.

1. Update the tool list.
2. Repeat device and extension reviews.
3. Review MCP servers and API keys.
4. Remove anything no longer used.
5. Update the data rules for new tools.
6. Remind the team how to report a problem.

Frequently asked questions

Do small teams need a formal AI policy?

You need written rules, but they can be short. One page on approved tools, data limits and reporting is enough for many teams. Keep it where people will find it, and update it when tools change.

Should we ban AI tools until we have a policy?

That often pushes use onto personal accounts, where you have less control. It usually works better to approve a few tools quickly and set simple data rules while you finish the rest.

How often should we check AI permissions?

Quarterly is a sensible rhythm for small teams, plus a check when you set up a device, add a new AI tool, or someone leaves.