Cybersecurity basics to fix before adopting AI
By the InTheory team · Updated November 2026
AI tools make whatever access and data your staff already have easier to find and move. Before rolling them out, fix the basics: multi-factor authentication everywhere, clean user access, tightened file sharing, clear rules on what data goes into AI, tested backups, and checks on AI vendors' data handling.
- AI assistants surface what users can already reach, so oversharing becomes visible.
- CISA's core steps include multi-factor authentication, strong passwords, updates and phishing awareness.
- Write simple rules on what data may and may not go into AI tools.
- Regulated firms should map AI use against HIPAA, the FTC Safeguards Rule or professional duties with their advisers.
Why AI raises the stakes
Most AI tools for business do not create new security holes on their own. What they do is make existing ones easier to reach. An assistant connected to your email and files can find, summarize and move information much faster than a person clicking through folders. If a junior employee can technically open the partners' compensation spreadsheet because of an old sharing link, an AI assistant may surface it in answer to an innocent question.
AI also adds new places for data to go. Staff paste client details into a free chatbot, connect a meeting recorder to their calendar, or give an automation tool access to the CRM. Each is a new vendor holding your data. Fixing the basics first means AI makes you faster without making you more exposed.
Start with CISA's core steps
The U.S. Cybersecurity and Infrastructure Security Agency recommends four core steps for everyone: recognize and report phishing, use strong passwords, use multi-factor authentication and a password manager, and keep software updated. For organizations it adds steps such as logging, backups, encryption and an incident response plan. These are not glamorous, but they stop a large share of everyday attacks.
In practice for a small firm, that means multi-factor authentication on every account that supports it, starting with email, Microsoft 365 or Google Workspace, your accounting system and any system holding client data. Use a company password manager so staff do not reuse passwords. Turn on automatic updates for devices and browsers. Run short phishing awareness training, because AI is also making scam emails more convincing.
If you only do one thing this month, turn on multi-factor authentication for email and every system that holds client data, and confirm it for every user, including the owner and any shared accounts. It is inexpensive, usually built into tools you already have, and it blocks a large share of account takeovers.
Clean up who has access to what
This review is worth doing on a schedule, such as every quarter, and every time someone leaves.
- Remove accounts for former employees and contractors in every system, not only email.
- Review admin rights and limit them to the few people who need them.
- Check shared mailboxes and service accounts, and confirm who controls them.
- Make sure every important system has at least two company admins, so you are not locked out if one leaves.
- Record who owns each system in a simple register.
- Confirm that consultants and vendors use their own named accounts, not shared logins.
Tighten file sharing
Before turning on an AI assistant that reads your files, review how documents are shared. In Microsoft 365, check SharePoint sites, Teams channels and OneDrive folders for links shared with everyone in the organization or anyone with the link. In Google Workspace, review shared drives and files shared outside the company. Prioritize folders holding HR records, payroll, client financial data, health information and legal documents.
The goal is simple: people should be able to open what they need for their job and not much more. This is often the biggest piece of preparation, and your managed IT provider can usually help. It also reduces risk from ordinary mistakes, such as forwarding the wrong file to a client.
After the cleanup, test it. Ask a non-admin employee to search for a few sensitive terms in SharePoint or Google Drive and confirm nothing appears that should not. It is a quick, practical check that the permission changes worked.
Set rules for data and AI tools
Write a short AI usage policy in plain language. List the approved tools and the accounts staff should use, which should be company business accounts rather than personal free versions. Say which data must never be entered into AI tools, such as social security numbers, bank details, passwords, health information or privileged communications, unless a specific tool has been approved for that purpose. Require a person to review anything AI produces before it goes to a client.
Ask each AI vendor how it handles your data: whether it trains models on it, where it is stored, how long it is kept, and who can access it. Prefer tools that let you turn off training on your data and that offer admin controls. Keep a list of approved tools and review it as staff request new ones.
Give staff an easy way to ask. A short list of approved tools in a shared document, plus a named person who answers questions within a day, prevents most shadow AI use. People usually reach for unapproved tools because they do not know an approved one exists.
Know your regulatory obligations
Some firms have specific duties that apply to AI as much as any other technology. Health practices handling patient information need to consider HIPAA and whether vendors will sign business associate agreements. Tax preparers and many financial firms are covered by the FTC Safeguards Rule, which requires a written information security program, and the IRS publishes guidance for tax professionals in Publication 4557. Law firms have confidentiality duties, and ABA Formal Opinion 512 addresses lawyers' use of generative AI. Colorado's AI Act adds requirements for certain high-risk uses of AI.
None of this is legal advice. The practical step is to ask your counsel or compliance adviser to review your AI policy and planned tools before rollout, and to keep a record of what was approved.
Back up and plan for incidents
Make sure critical data is backed up, that backups are stored separately from your main systems, and that someone has actually tested restoring from them. Cloud services do not always back up your data the way people assume, so check what your Microsoft 365, Google Workspace and line-of-business vendors provide and whether you need more.
Write a one-page incident plan: who to call if an account is compromised or data is exposed, how to reach your IT provider and cyber insurer out of hours, and who decides on notifying clients. Many cyber insurance policies require specific controls, such as multi-factor authentication, so check your policy too. InTheory would typically confirm these basics with a firm's IT provider before scoping any AI work, because the cheapest security problem is the one fixed before rollout.
Common questions
Do I need to fix security before using ChatGPT or Copilot?
You should at least cover the basics: multi-factor authentication, company accounts rather than personal ones, rules on what data may be entered, and a review of file sharing if the tool can read your files. These steps reduce the most likely problems and are worth doing regardless of AI.
What is the biggest AI security risk for small businesses?
Usually it is data going where it should not: staff pasting confidential information into unapproved tools, or AI assistants surfacing files that were shared too broadly. Clear rules, approved business accounts and tidy permissions address most of it.
Who should own cybersecurity in a small firm?
A named person inside the business should be accountable, even if a managed IT provider does most of the work. That person makes sure access reviews happen, policies are current and the incident plan is known, and they coordinate with advisers on regulatory questions.