Products SEO & GEO Agent Paid Ads Agent Website Agent Conversion Agent
How it works Pricing FAQ Blog Book a call

AI Security Risks: What a Small Business Actually Faces

Updated · 6 min read

Short answer A small business faces three AI security risks: customer or company data pasted into AI tools that may keep or train on it, scammers using AI to fake voices, videos and emails, and AI agents with access to your accounts making wrong moves or being tricked. Most fixes are cheap: business plans, call-back rules and limited access.

Want help applying this to your business? Talk it through on a free 30-minute call.

Book a call

Which AI security risks actually reach a small business?

Three. Most of what you read about AI security (stolen models, poisoned training data) is a problem for the companies that build AI, not for a dental office or a roofer that uses it.

RiskHow it happensFirst fix
Data exposureSomeone pastes customer details, contracts or passwords into a free AI chatBusiness plan with training off, plus a "never paste" list
AI-made scamsA cloned voice, fake video call or flawless email asks staff to pay or share a loginCall back on a number you already have before any payment
Agent mistakesAn AI tool with access to your site or ad account does the wrong thing, or obeys hidden instructionsLowest access level, spending caps, your approval for big changes

The order matters. Data exposure happens quietly every day. Scams cost the most money. Agent mistakes are newer and grow with how much access you hand out.

How big is the threat?

Big for large companies, and the scam part reaches everyone. IBM's 2026 study puts the global average cost of a data breach at $4.99 million, "a 12% increase over last year and a record high," and reports a 56% rise in AI-driven attacks, "led by AI deepfake impersonations and AI-enabled malware" (IBM: Cost of a Data Breach Report 2026). That average comes from larger organizations; a small firm's breach costs less but can still sink it.

The FBI's numbers are closer to home. In 2025 its Internet Crime Complaint Center logged more than 22,000 complaints mentioning AI, with losses over $893 million, and business email compromise (a fake email from a boss or supplier asking for money) cost over $3 billion across all reports (FBI IC3: 2025 Internet Crime Report). The FBI also notes that only part of that fraud uses AI: businesses reported over $30 million lost to email scams involving AI.

Is my business data safe in ChatGPT and other AI tools?

It depends on the plan. On OpenAI's consumer plans, your chats can become training material unless you switch it off:

"When you use our services for individuals (Free, Plus, Pro), we may use your content to train our models." (OpenAI Help Center)

The opt-out is under Settings > Data controls: turn off "Improve the model for everyone." Business plans work the other way: "By default, we don't use inputs or outputs from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu, or our API to improve our models" (OpenAI Help Center; OpenAI: Enterprise privacy).

Training is only one way data leaks. A tool can also be breached, or store chats in another country (see Chinese AI models). Canada's cyber agency gives the plain rule:

"Avoid providing PII or sensitive corporate data as part of the queries or prompts." (Canadian Centre for Cyber Security)

PII means personally identifiable information: names, addresses, phone numbers, health or account details. For patient data in the US there is a separate rule set; see Is ChatGPT HIPAA compliant?

The biggest gap is usually staff using free tools nobody approved, often called shadow AI. A one-page list of approved tools and a "never paste" list closes most of it.

How do criminals use AI against small businesses?

They use it to impersonate people you trust. The FBI warned that criminals use AI to "generate short audio clips containing a loved one's voice" and to "generate videos for real time video chats with alleged company executives" (FBI IC3, December 2024). The same warning says AI lets them write believable messages faster and to more people, so the old tells (bad spelling, odd wording) are gone.

For a small business the usual targets are:

  • Payments. A "supplier" emails new bank details, or the "owner" calls the bookkeeper asking for an urgent wire.
  • Logins. A convincing message asks staff to "confirm" an ad account, website or email password.
  • Reputation. Fake reviews or fake videos about your business.

The fix is a process, not software. Agree on a rule that any new payment detail or urgent money request gets checked by calling back a number you already have, never one in the message. The FBI suggests families agree on a secret word or phrase; an office can do the same. Turn on multifactor authentication (a code or app prompt on top of the password) for email, banking, ad and website accounts; CISA says it "can make you much more secure" (CISA: More than a Password).

What can go wrong if I let an AI agent into my accounts?

An AI agent is software that takes actions on its own, like editing your website or changing ad budgets (see agentic AI). Two things can go wrong.

It makes a bad call. The security group OWASP calls this "excessive agency" and names three causes: "excessive functionality; excessive permissions; excessive autonomy" (OWASP: LLM06:2025). In plain words: it can do too much, reach too much, or act without asking.

It gets tricked. Prompt injection means instructions hidden in content the AI reads. OWASP explains that indirect prompt injection happens "when an LLM accepts input from external sources, such as websites or files" (OWASP: LLM01:2025). A web page, review or email could carry text that tells your agent to send data elsewhere. Defenses are improving but not complete. Anthropic, which makes Claude, wrote that even a 1% attack success rate "still represents meaningful risk" (Anthropic, November 2025).

OWASP's fixes translate directly for an owner:

  1. Give the tool its own login at the lowest access level that still does the job, never your owner login.
  2. Cap what it can spend at the platform level (daily budget, account spending limit), not only inside the tool.
  3. Approve big moves yourself. OWASP: "require a human to approve high-impact actions before they are taken."
  4. Check the change history weekly, and remove access for tools you stopped using.

Which jobs to hand over at all is covered in what to automate and what humans own.

What should I ask an AI vendor before giving it access?

Five questions cover most of it:

  • Is my data used to train models, and can I turn that off in writing?
  • Which AI model do you use, and where is my data stored?
  • What exactly can the tool change, and can I limit it?
  • Which actions need my approval, and can I set that list?
  • Can I see a log of every change, and undo it?

A vague answer to any of these is your answer. For a fuller checklist when choosing a marketing agent, see AI SEO agent review: 8 checks.

Example

A bookkeeper gets a call from "the owner"

The voice sounds right and asks for a $9,000 transfer to a new vendor today. The office rule says every new payee gets a call back on the owner's saved number. The real owner never called. Nothing is lost.

A clinic receptionist uses free ChatGPT

She pastes patient emails to draft replies. Training is on, and the data is health information. The clinic moves to a business plan, removes names before pasting, and keeps patient replies out of AI tools until its compliance check is done.

A contractor connects an AI tool to his ad account

He gives it standard access, not admin, and sets a $50 daily budget cap in the ad platform. The tool suggests doubling spend on one campaign and waits for his approval. If it misfires, the cap keeps the damage small.

FAQ

Can AI tools steal my passwords?

Rarely by itself. The risk is staff pasting a password into a chat, or an AI-written phishing message luring them to a fake login page.

Is a paid AI plan safer than the free one?

For training, yes: OpenAI's business plans don't train on your data by default. Paid personal plans (Plus, Pro) can, like Free, unless you opt out.

Should a small business ban AI tools?

A ban usually pushes use out of sight. Approving a few tools on business plans, with clear rules on what not to paste, works better.

Do I need special AI security software?

Rarely at small-business size. Multifactor login, limited access, call-back rules for payments and a short written policy cover the main risks.

Different for you?

Keep learning

NextWhite Hat AI: How Ethical Hacking Tools Protect a BusinessWhat white hat AI really does for a small business: AI scanners that find holes, AI threat detection on laptops, free scans, and the permission rule.

Ask us one question

One line is enough. We reply by email.

Sources

  1. IBM: Cost of a Data Breach Report 2026
  2. FBI IC3: 2025 Internet Crime Report
  3. FBI IC3: Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud (PSA I-120324-PSA, December 3, 2024)
  4. OpenAI Help Center: How your data is used to improve model performance
  5. OpenAI: Enterprise privacy
  6. Canadian Centre for Cyber Security: Generative artificial intelligence (ITSAP.00.041)
  7. OWASP: LLM01:2025 Prompt Injection
  8. OWASP: LLM06:2025 Excessive Agency
  9. Anthropic: Mitigating the risk of prompt injections in browser use (November 2025)
  10. CISA: More than a Password (multifactor authentication)