AI security risks became headline news when OpenAI, the company behind ChatGPT, confirmed it suffered a security breach. For SMB owners already questioning whether to adopt AI tools or trying to control employee use of them, this incident offers five critical lessons about what can go wrong and what you need to do differently.
What happened in the OpenAI breach?
A hacker gained access to OpenAI’s internal systems through what security researchers describe as vulnerabilities tied to the company’s AI training infrastructure. While OpenAI stated that customer data was not accessed, the breach exposed internal discussions and technical details about how the company builds its models.
The incident matters because it happened to one of the most well-funded, security-conscious AI companies in the world. If OpenAI can be compromised, the dozens of smaller AI vendors your employees might be using (often without IT approval) face even greater risk.
For an SMB, this translates directly: when your team uploads customer lists, financial projections, or proprietary processes into an AI tool, you are trusting that vendor’s security posture. Most business owners have no visibility into whether that trust is justified.
Why do AI companies face unique AI security risks?
Traditional software companies protect a static product. AI companies, by contrast, are constantly ingesting new data to train and refine models. That creates a much larger attack surface.
Training data often includes vast datasets scraped from the internet, uploaded by users, or licensed from third parties. Each source is a potential entry point. The systems that store and process this data are under intense pressure to scale quickly, which can lead to security shortcuts.
The OpenAI incident reportedly stemmed from these training and development practices. Attackers exploited the complexity and speed of AI development, areas where governance often lags behind innovation.
For your business, this means asking a hard question: do you know how the AI tools your employees use handle the data they collect? Most SMBs do not. If an employee pastes a client contract into ChatGPT or a similar tool to summarize it, where does that text go? Who can see it? How long is it retained?
Without answers, you are operating on hope, not control.
What data is actually at risk when employees use AI tools?
The risk is not theoretical. We have seen SMBs inadvertently expose:
- Customer names, email addresses, and purchase histories uploaded to AI marketing tools.
- Financial statements and forecasts entered into AI-powered spreadsheet assistants.
- Employee performance reviews and compensation data pasted into AI writing tools.
- Proprietary manufacturing processes or service workflows described to AI chatbots for process optimization.
- Legal documents and contracts summarized by generative AI without attorney review of the tool’s terms of service.
Each of these actions can trigger compliance violations. If you are subject to HIPAA (Health Insurance Portability and Accountability Act), entering patient information into an unapproved AI tool is a breach. If you handle credit card data under PCI DSS (Payment Card Industry Data Security Standard), the same applies. Even if you are not in a regulated industry, exposing customer data can violate your own privacy policy or contractual obligations.
The OpenAI breach is a reminder that once data leaves your environment, your ability to protect it depends entirely on someone else’s defenses.
How should SMBs govern AI tool adoption?
The answer is not to ban AI. Prohibition does not work. Employees will use the tools anyway, just without telling you. That creates shadow IT, which is far more dangerous than transparent, governed use.
Instead, establish an AI approval process. Before any employee uses a new AI tool for business purposes, require them to submit it for review. Your IT partner (or internal IT lead) should evaluate:
- Where the vendor stores data (U.S.-based servers are generally preferable for compliance reasons).
- Whether the vendor uses your inputs to train models available to other customers (many do by default, unless you pay for enterprise tiers).
- What the vendor’s breach notification and incident response policies are.
- Whether the tool integrates with other systems and what permissions it requests.
- Whether a Business Associate Agreement (BAA) or Data Processing Agreement (DPA) is available if you handle regulated data.
This does not need to be a multi-week audit. A simple checklist and a 15-minute conversation can surface most red flags. The goal is accountability, not perfection.
What belongs in an employee AI policy?
A policy does not have to be long, but it must be clear. Employees need to know:
- Which AI tools are pre-approved for business use (and which are not).
- What types of information may never be entered into any AI system (e.g., Social Security numbers, credit card numbers, patient health information, attorney-client privileged material).
- Who to ask before adopting a new tool.
- That use of unapproved tools will be treated as a policy violation, with consequences.
Post the policy where people will see it. Include it in onboarding. Reference it in your acceptable use policy. And when someone asks to use a new tool, respond quickly. If the approval process is a black hole, employees will route around it.
One manufacturing client we work with created a simple internal wiki page listing approved AI tools, each with a one-sentence description of acceptable use. When an employee wants to try something new, they submit a request via a shared Slack channel. The IT lead and owner review it within 48 hours. Approvals are added to the wiki. It is lightweight, fast, and effective.
Do SMBs need to worry about AI security risks if they are not using AI yet?
Yes, because your employees almost certainly are. Studies show that a significant percentage of knowledge workers use generative AI tools without informing their employer. They are solving real problems (drafting emails, summarizing reports, generating ideas), but they are doing so without oversight.
If you have not had a conversation with your team about AI use, you are not choosing to avoid AI security risks. You are choosing to ignore them. The risks exist whether you acknowledge them or not.
The first step is a simple survey or discussion: “Who is currently using AI tools for work? Which ones? For what tasks?” The answers will often surprise you. Once you know what is happening, you can decide what to approve, what to redirect to safer alternatives, and what to prohibit.
This is not about control for its own sake. It is about protecting the business you have built. A data breach, a compliance penalty, or a lost client because proprietary information leaked is not an abstract risk. It is a business-ending event for many SMBs.
What should SMBs do right now?
Three actions will address the majority of AI security risks for most SMBs:
First, create or update your acceptable use policy to address AI. Make it clear that business data may not be entered into unapproved tools. Define what “business data” means (hint: if you would not want it posted on your website, it is business data).
Second, inventory what AI tools are already in use. Ask your team directly. Check browser histories if you manage devices (and if your employee handbook permits it). Look at software subscriptions and credit card statements for SaaS charges you do not recognize.
Third, establish a lightweight approval process. One person (your IT lead, your MSP, or a trusted manager) should be the gatekeeper. Employees submit requests, the gatekeeper evaluates them using a standard checklist, and decisions are documented.
None of this requires a consultant or a six-month project. It requires an afternoon to draft the policy, a team meeting to communicate it, and a commitment to enforce it consistently.
If you need help with the technical evaluation (for instance, understanding whether a tool’s terms of service expose you to liability), that is where a partner experienced in AI adoption security risks can save you time and reduce exposure.
How does this fit into a broader IT strategy?
AI governance is not a separate initiative. It is part of your overall approach to data security, vendor management, and compliance.
If you already have a process for approving new software vendors, extend it to cover AI tools. If you already require Data Processing Agreements for vendors who handle customer information, require them for AI vendors too. If you already conduct periodic access reviews (who has access to what), include AI tool accounts in that review.
The principles do not change. What changes is the speed at which new tools appear and the ease with which employees can adopt them. A SaaS subscription used to require a purchase order and a contract. Now it requires a credit card and 60 seconds. Your governance has to match that pace, which means lightweight processes that can be executed quickly.
For SMBs without a formal IT strategy, AI adoption is often the forcing function that makes one necessary. You cannot govern AI in isolation. You need a framework for evaluating all technology decisions, and AI is simply the most urgent example. We help clients build that framework, starting with practical services and solutions that match their size and risk profile.
FAQ: AI security risks for SMBs
Can I just block access to ChatGPT and other AI tools on my network?
You can, but it is not an effective long-term strategy. Employees will use AI tools on personal devices, home networks, or mobile hotspots. Blocking creates shadow IT and eliminates your visibility. A better approach is to approve specific AI tools, provide guidance on safe use, and monitor for unapproved access.
How do I know if an AI vendor is secure enough for my business?
Start by reviewing the vendor’s security documentation (usually available on their website under Trust or Security pages). Look for SOC 2 Type II certification, which indicates independent audit of security controls. Ask whether they use your data to train models (if yes, can you opt out?). Check where data is stored and whether they will sign a Data Processing Agreement. If the vendor cannot or will not answer these questions, that is your answer.
What happens if an employee accidentally shares confidential information with an AI tool?
It depends on the tool and the data. Some AI platforms allow you to delete inputs and request removal from training data, though there is no guarantee of complete removal. If the data is regulated (HIPAA, PCI, etc.), you may have a breach notification obligation. Document what happened, assess the risk, consult with legal counsel if needed, and use the incident to reinforce your policy. Prevention is far easier than remediation.
Do I need a separate AI policy or can I add AI to my existing acceptable use policy?
Either approach works. The key is clarity. If your acceptable use policy is already comprehensive and well-understood, adding an AI section makes sense. If employees rarely read it, a standalone AI policy (one page, plain language) might get more attention. The format matters less than the content and the communication.
Are there AI tools that are safer for SMBs to use than others?
Yes. Enterprise tiers of major AI platforms (Microsoft Copilot for Microsoft 365, Google Workspace AI features, etc.) typically include stronger data protection, Business Associate Agreements for HIPAA, and commitments not to use your data for model training. They cost more, but the additional protection is worth it for business use. Free consumer versions of AI tools should be treated as public forums: assume anything you enter could become visible to others.
How often should we review our approved AI tools list?
Quarterly is a good baseline. AI tools change rapidly (new features, new privacy terms, new security incidents). Schedule a brief review every 90 days to confirm that approved tools still meet your criteria and to evaluate any new requests. If a major security incident occurs (like the OpenAI breach), that is a trigger to review immediately, even if you are mid-cycle.
Keep reading
Sources
Source: OpenAI hacking incident exposes mounting risks in AI arms race – Financial Times