What AI Security Means for a Small Business

TL;DR: AI security for a small business has three parts: keeping your data safe when staff use AI tools, securing any AI you build or connect to your systems, and defending against criminals who now use AI for phishing and voice fraud. Most of the first steps are habits and settings you control, not new software.

AI security for a small business sounds like a job for a large company's security team. It is closer to the security you already practice, applied to a new tool. The questions owners ask me are practical: what is it, what should my staff never type into a chatbot, and is our data safe?

The honest answer: AI security depends on the way you set it up and use it, and that is mostly in your hands.

What is AI security?

Start with the federal definition. NIST's AI Risk Management Framework says AI systems "that can maintain confidentiality, integrity, and availability through protection mechanisms that prevent unauthorized access and use may be said to be secure." (NIST AI 100-1, January 2023) In plain words, AI security means the right people see the right data, the system does what you expect, and it keeps working.

NIST's later profile for generative AI, the kind behind chatbots and writing assistants, adds the part that makes AI security new. It says these systems "present two primary information security risks." They can lower the barrier for attackers, and "simultaneously, it expands the available attack surface, as GAI itself is vulnerable to attacks like prompt injection or data poisoning." The same profile lists data privacy, meaning leakage and unauthorized disclosure of personal or sensitive data, as its own risk. (NIST AI 600-1, July 2024)

For a small business, I find it clearest to sort AI security into three parts:

1. Your data in other people's AI. What happens to customer records, contracts and numbers when staff paste them into AI tools.
2. AI you build or connect. A chatbot on your website, an assistant linked to your inbox, an automation that reads invoices. These become part of your systems and need the same care.
3. AI aimed at you. Phishing emails written by AI, cloned voices on the phone, fake executives on a video call.

In my view, most small businesses need to go deep on the first and third parts of AI security. The second matters as soon as AI touches real data.

Is AI safe for my small business data?

It can be, and the deciding factor in AI security is usually not the AI itself. It is which account your staff use and what they put into it.

Use is already widespread. Verizon's 2026 Data Breach Investigations Report, drawing on data loss prevention data from the organizations it studies, found that 45% of employees are now regular users of AI on their corporate devices, up from 15% the year before, where regular means at least once every 15 days. And 67% of the people reaching AI services from work devices did it through non-corporate accounts, down slightly from 72% the year before. (Verizon 2026 DBIR)

Put those side by side and the AI security gap is plain. Almost half of employees use AI at work, and two in three reach it through accounts the business does not control, on terms the business never chose.

OWASP, the nonprofit behind widely used security risk lists, ranks "Sensitive Information Disclosure" second in its 2025 Top 10 for AI language model applications. Its advice is direct: users "need to understand the risks of unintentionally providing sensitive data, which may later be disclosed," and organizations should "provide guidance on avoiding the input of sensitive information." (OWASP LLM02:2025)

So good AI security looks like this. The business picks the AI tools and holds the accounts. Someone has read what those plans say about training, retention and who can see chats. Brian compared the major AI companies' own terms in whether staff can put customer data into AI tools (opens in a new tab). Staff know what stays out. A ban rarely works as AI security, because people keep using tools that save them time. I have written about why enabling approved AI works better than banning it (opens in a new tab).

What should you never paste into an AI tool?

The question people search is "what to never say to AI." For AI security in a business, the useful version is what never goes into a prompt, a file upload or a chat, unless the tool is one your business has approved for that kind of data.

• Passwords, keys and access codes. Login details, API keys, Wi-Fi passwords, door codes, recovery codes. OWASP lists security credentials among the sensitive information an AI system can expose.
• Personal information about customers or staff. Social Security numbers, driver's license numbers, bank account and card numbers. In Massachusetts, a resident's name combined with any of those is personal information under the state's data security regulation, 201 CMR 17.00.
• Health information. Patient notes, diagnoses, insurance details. If HIPAA covers your business, its rules do not pause for AI tools.
• Financial details that are not public. Bank statements, payroll, unreleased results, pricing a customer negotiated in confidence.
• Contracts and legal documents. Especially anything under a confidentiality clause, which may forbid sharing it with a third party at all.
• Code and system details. Source code, server addresses, network diagrams, security settings.
• Anything a client asked you to keep confidential.

One AI security rule from federal guidance is worth adopting. A joint information sheet from CISA, the NSA, the FBI and partner agencies in Australia, New Zealand and the UK says "the output of AI systems should be classified at the same level as the input data." (CISA, NSA, FBI and partners, AI Data Security, May 2025) If a confidential spreadsheet goes in, the summary that comes out is confidential too.

What can staff paste? Public information, marketing drafts, general questions, and documents with names and numbers removed. Swapping "Jane Smith, account 4417" for "a customer" often costs little in the answer's quality.

What changes when you build or connect AI?

Once an AI tool can read your inbox, answer customers on your website or pull records from your systems, it becomes part of your systems, with AI security risks a plain chatbot does not have.

OWASP's 2025 Top 10 for these applications is the clearest public map. (OWASP Top 10 for LLM Applications 2025) For small business AI security, five of the ten deserve attention first:

• Prompt injection. Instructions hidden in the content the AI reads. OWASP describes the indirect version as what happens "when an LLM accepts input from external sources, such as websites or files." An email or web page can carry text that tries to steer your assistant. Qurtoo, our free prompt engineering resource, has a short explainer on prompt injection (opens in a new tab).
• Sensitive information disclosure. The AI repeats something it was given to someone who should not see it, such as one customer's details in another customer's chat.
• Supply chain. The models, plugins and connectors you add are software from other companies, with their own weaknesses.
• Improper output handling. AI output passed straight into another system, an email or a database without checks.
• Excessive agency. The AI can do more than its job needs: send, delete, pay or share.

NIST notes that researchers "have already demonstrated how indirect prompt injections can exploit vulnerabilities by stealing proprietary data or running malicious code remotely on a machine." (NIST AI 600-1) That is why the AI security controls that matter most sit outside the AI: what data it can reach, what actions it can take, and which of those need a person to approve.

The joint AI data security guidance lists ten AI security best practices, from "Classify data and use access controls" to "Delete data securely," and most are habits that already protect any business system. Our knowledge base covers agentic security and AI agent governance for a small business in more depth.

Before you connect anything, ask three AI security questions: what can it read, what can it do without asking, and who finds out when it gets something wrong? If nobody can answer, it is not ready for customer data.

How are criminals using AI against small businesses?

Mostly to make old scams more convincing. The FBI's Internet Crime Complaint Center received 22,364 complaints in 2025 that referenced AI, with adjusted losses of $893,346,472. (FBI IC3, 2025 Internet Crime Report)

For businesses, the key line is business email compromise: a criminal poses as an executive or vendor to redirect a payment. The IC3 report says chat generators "can quickly create official-sounding emails mimicking a company's CEO or other officials," and that "voice cloning can also be used to request wire payment." Businesses reported over $30 million in such losses involving AI in 2025, out of $3,046,598,558 in total business email compromise losses.

That gap is worth reading carefully. The AI-linked losses are about 1% of the total, and the FBI notes that "not all are AI-enabled." AI makes the message smoother. The scam is the same request to change where money goes.

The video version has already happened at scale. In 2024, the engineering firm Arup confirmed that "fake voices and images were used" in a scam where a finance employee in its Hong Kong office joined a video call with deepfake re-creations of the chief financial officer and other staff, then sent about $25.6 million across 15 transactions. (CNN, May 2024)

Verizon adds perspective. AI-assisted text in malicious emails has grown, yet phishing's share as the way attackers first get in "has barely moved over the past few years," which the report reads as a sign that AI may not be raising the success rate against the organizations it studies. (Verizon 2026 DBIR) For AI security, older defenses still work, as long as they do not depend on spotting typos.

The FBI recommends verifying a caller's identity "by hanging up the phone" and calling back, and creating "a secret word or phrase" with family. (FBI PSA, December 2024) The FTC gives the same advice for cloned voices: "Use a phone number you know is theirs." (FTC, March 2023) For a business, that becomes three AI security rules:

1. Payment changes are confirmed by calling a number you already have on file, never one from the email or call asking for the change.
2. Any new payee or wire above a set amount needs a second person to approve it.
3. Urgency and secrecy are the signal. "Keep this between us" and "it has to go today" are reasons to slow down, not speed up.

Voice fraud has its own patterns, and I wrote a plain guide to vishing and AI voice clones (opens in a new tab) for teams that answer the phone.

What does AI security mean for a Massachusetts small business?

Massachusetts has been explicit that existing law covers AI. On April 16, 2024, Attorney General Andrea Joy Campbell issued an advisory saying the state's consumer protection, anti-discrimination and data security laws apply to AI. It says AI developers, suppliers and users "must take the necessary and appropriate steps to safeguard personal information used by those systems," pointing to 201 CMR 17.03 and 17.04, and "are expected to comply with the breach notification requirements set forth in the statute." (Massachusetts Attorney General, AI Advisory)

For a business in Worcester, Shrewsbury or anywhere in the state, that means your written information security program should cover AI security, because staff using AI tools with customer data is part of how that data is handled. The advisory also names deepfakes and voice cloning used to deceive people among the risks it addresses.

The federal picture on AI security matches. When the FTC announced its Operation AI Comply enforcement sweep, then Chair Lina Khan said "there is no AI exemption from the laws on the books." (FTC, September 2024)

The fraud side is local too. People in Massachusetts filed 22,936 complaints with the FBI's IC3 in 2025, reporting $410,924,066 in losses, which put the state 13th in complaints and 14th in losses nationally. (FBI IC3, 2025 Internet Crime Report) To read the state rules at their source, MyRHC keeps an index of Massachusetts compliance sources (opens in a new tab).

Where should a small business start with AI security?

None of these AI security steps needs new software.

1. List the AI tools in use. Ask the team which ones they use and with which account. Expect more than you think, and make it safe to answer honestly.
2. Pick the approved tools and hold the accounts. Choose business accounts the company controls, and read what each plan says about training, retention and who can see conversations. Brian priced five business AI plans side by side (opens in a new tab).
3. Write a one-page AI security rule. What never goes into AI (the list above), what may, and who to ask. Short enough that people read it.
4. Add AI to your written security program. Name the approved tools and the data rules. In Massachusetts, 201 CMR 17.00 already requires a written program from any business that holds residents' personal information, and AI security belongs in it.
5. Lock the payment process. Callback to a known number, two people on new payees and large transfers, and permission to slow down.
6. Before connecting AI to real data, map what it can reach. Give it only the access the job needs, keep a person on actions that send, pay, delete or share, and keep a log someone actually reads.
7. Use an AI security framework when you outgrow a checklist. NIST's AI Risk Management Framework is free and "intended for voluntary use," built around four functions: Govern, Map, Measure and Manage. (NIST AI RMF)

Good AI security does not require a security department. It requires someone to own it: decide which tools, write it down, and check back a month later.

This article is general information, not legal advice. Kief Studio is not a law firm, and reading it does not create an attorney-client relationship. Laws and regulations change, so check the current text and talk with a licensed Massachusetts attorney about your situation.