ESET’s Tony Anscombe on Shadow AI and Cybersecurity Risks for Small Businesses

Artificial intelligence is rapidly becoming part of everyday business operations, helping small businesses improve productivity, automate tasks, and accomplish more with fewer resources. However, as employees increasingly turn to AI tools, many businesses are discovering that the technology introduces security risks they may not fully understand or be prepared to manage.

One growing concern is “shadow AI,” the use of artificial intelligence applications without an employer’s knowledge or approval. From employees uploading sensitive customer information into public AI chatbots to AI agents accessing business systems, these activities can expose companies to data breaches, privacy violations, and other cybersecurity threats.

The challenge is particularly significant for smaller businesses that may lack dedicated IT teams or cybersecurity specialists. According to an ESET small business cyber readiness survey discussed during the interview, 81% of U.S. small and midsize business respondents said they were integrating AI into their operations, while 75% acknowledged that AI introduces new risks.

To explore these challenges and what small business owners can do about them, I recently sat down with Tony Anscombe, Chief Security Evangelist at ESET, for another conversation about the evolving cybersecurity landscape.

During our discussion, Tony explained the dangers of shadow AI, what businesses should consider before adopting AI agents and downloadable AI skills, and how owners can establish practical security policies without sacrificing productivity. We also examined ESET’s cybersecurity research, including the disconnect between businesses’ confidence in their security and the number of incidents they actually experience.

Most importantly, Tony shared actionable recommendations that small business owners can implement to better protect their employees, customers, and business information as AI becomes an increasingly important part of their operations.

Here is our conversation.

Leland McFarland: Hello, small business owners. My name is Leland McFarland, and I got a special episode here for you. I have Tony Anscombe, who is the Chief Security Evangelist for ESET. Hey Tony, glad to have you on.

Tony Anscombe: Hey, good to be back here again, Leland, and talking to you.

Leland McFarland: All right, perfect. Well, let’s dive right in. So Tony, in our last conversation, we discussed how AI is changing the cybersecurity landscape for small businesses. One particular concern that’s getting more attention is employees using AI applications without their company’s knowledge or approval. So let’s start with the basics: what is shadow AI, and why should small business owners care about it?

Tony Anscombe: Well, I think you summed it up actually pretty well there of using something without the company’s approval. So, i.e., this is where you want to have the benefit of using AI in the workplace, but the employer hasn’t actually turned and said, “If you’re going to use a chatbot or whatever it might be, or you’re going to use AI to research something, then this is the one that you’re going to use, and you’re going to use a logged-in version, so i.e., it’s within the company confines and security.” Instead of that, they’ve said nothing, and every employee goes off and uses whatever they want.

Now, this is not the first time we’ve seen something like this, by the way, is it? Because if we go back a few years—and I hate the word “pandemic” because everything gets related back to that word—but if we remember back, at the start of the pandemic, lots of people used file sharing and lots of people used different conferencing systems, and companies had to quickly turn and say, “Okay, this file sharing, this conferencing system.” We’re kind of at the same point. Companies that haven’t turned and said “this AI” or “within these guardrails,” then they’re kind of opening the Pandora’s box that people will use shadow AI, i.e., unauthorized applications. You don’t really know where the data is going, you don’t know what’s providing the answers, and employees will go and do it.

And it’s even a problem even if you do have an approved version, because if your approved version is Copilot, but somebody likes using Gemini, then they might go off and use Gemini still, even on a personal device like a BYOD off their phone. So that’s what shadow AI is, is those unauthorized applications.

Leland McFarland: Right. So AI tools can make small teams much more productive, whether it’s helping with customer service, creating marketing content, managing documents, or automating administrative tasks. But with employees increasingly relying on these applications, where can that everyday use start to create security or privacy risks for a business?

Tony Anscombe: Well, now imagine our own jobs, Leland. I don’t know how much you use AI—in fact, I’m just checking you’re not AI.

(Leland laughs)

No, no, no. But think about this: so you’re writing a presentation, or you’re writing a document, or even writing an email, and you ask AI to do something to the content, whether it’s improve the content of the email you’re going to send somebody. Suddenly, the content of that email is being uploaded into what could be a publicly trained LLM, and that data then becomes part of the future training of the LLM, especially if it’s in shadow AI. So suddenly you’ve got a privacy breach, a data leak—well, a data leak in that case. But what happens if there was customer information in there? Have you now got a privacy breach issue? So now with that, you then enter legal and compliance penalties.

And then you’ve got the output from that AI chat that’s then going to be sent to a customer or even sent internally. Who’s verifying? Who’s validating that it’s accurate? So if you’re sending it outside the customer, do you trust the output from the AI? Yeah, I wouldn’t. But people do, and unfortunately, that implicit trust causes business issues. But it’s the data leakage, the hallucinations, and the legal and compliance issues that really are causing businesses trouble with this.

Leland McFarland: Right. So, I think one of the challenges for employees is understanding what information is safe to share with AI. Something as simple as asking an AI tool to summarize an email or, as you mentioned before, or analyzing a spreadsheet could potentially expose sensitive information. What kinds of information should employees never put in a public, unapproved AI tool?

Tony Anscombe: Well, obviously, privacy information for me has to be kind of the top of that list. I mean, anything that’s got employee data, customer data, or personal information that could be used or abused in that way, then that needs to be the top of the list for me. But from a company perspective, it’s much broader than that.

What about financial data? What if somebody in your accounts department turns and says, “Can you show me the trend in our sales line and draw me a nice chart?” Suddenly they’re uploading the financials for the last 12 months into a public model; then you’re disclosing what your revenue is and maybe where your revenue is coming from. So any financial information.

Or, in fact, I think a much better example could be a contract. You get a business contract in that maybe somebody has sent you for signing, and maybe you ask AI to summarize the contract or check if there is anything unusual in the contract. But in uploading the contract, you’ve just disclosed that there is a contract and the details of that contract. So something like that is dangerous.

But then there’s also product information. What if you and I had this wonderful jam company where we’re making pots of jelly jam, and it’s got a unique recipe? You don’t want that recipe or the IP of the company to be uploaded, because suddenly you’ve lost your uniqueness.

So there’s all sorts of very sensitive company information that you might not want uploaded. And it may even extend to, for example, marketing information. If the marketing team are about to do a campaign around a new product that you’re releasing in three months’ time, and they ask it to create graphics, then that’s competitive information you wouldn’t want in the public market. So realistically, I think there’s an awful lot of information that could leak out of a company into an AI model that companies need to protect.

Leland McFarland: Right. So, ESET completed a small business cyber readiness survey earlier this year. It found that 81% of U.S. SMBs responded saying that they were integrating AI, while 75% acknowledged that AI introduces risks. For small business owners who want to better understand these risks, where should they go for reliable information and practical guidance?

Tony Anscombe: Well, there’s a number of places. Firstly, I promote our own sources for that: we produce a lot of information on WeLiveSecurity.com and our corporate blog on eset.com. So I’d highly recommend following those and subscribing where you can, because we send out a weekly email with the summary of that week as well.

But if somebody wanted further information beyond that, I’d recommend they look at one of the national organizations. So in the U.S., you’ve got NIST. For example, NIST produced an AI Risk Management Framework and produce a lot of information. And the nice part about that is—that might sound complicated to small businesses to start with when I called it an AI Risk Management Framework—it starts from the point of zero knowledge. So for a small business owner that’s coming into this thinking, “What do I need to do, and where do I get more information?”, from that zero-knowledge point of view, NIST is actually a good place for somebody to go look.

Leland McFarland: All right. And for those who are watching this, I’ll have links in the description to all of these resources that he just talked about.

All right, so many small businesses don’t have a dedicated IT or cybersecurity team, and the business owners may be the person making all of the technology decisions. What are a few simple guardrails an owner can put in place without making it harder for employees to use AI productively?

Tony Anscombe: Well, firstly, there’s technically prohibiting access, but let’s say this is the first thing that actually a small business owner should go off and do, and that’s to create a policy: a single one-page document that goes out to all employees saying what the purpose of the document is, why this is a risk, what is approved, and, importantly, what is not approved (i.e., what is prohibited), what tools can be and can’t be used, guidance on that cut-and-paste that we just talked about (somebody cutting and pasting in or uploading information or documents, what can be and can’t be).

So this one-pager allows an employee to understand the risk, what can and shouldn’t be done, and then also the review of the fact-checking. Don’t ever send anything to a customer without fact-checking it first if it came out of AI. So I think if a small business owner did a policy of those native things to start with, that would be a good place to start.

Leland McFarland: Right. So, ESET has recommended providing employees with approved AI tools rather than simply banning unapproved applications. For a small business choosing between free AI services and paid business or enterprise versions, what should the owner look for when evaluating data privacy, security controls, and how these services handle company information?

Tony Anscombe: Yeah, so there’s a number of things to look for, and most of the AI services now provide these in a subscribed method. So if as a company you start paying to use AI, you can suddenly start explicitly excluding your data from training any future model or use. So, i.e., in effect, your own data stays within your own organization’s use of AI, but you’re enjoying the AI using the public side of what it provides you back.

That also allows you to then fully own the intellectual property of what you’re uploading, what you’re doing, and what’s generated. So you’re siloing your interaction with AI. To me, that’s probably the most important thing.

Then you get into centralized management, so i.e., I can see what people are interacting with, how they’re doing it, etc., but as long as I’m protecting my data, that’s the primary one.

And then there’s data deletion: how long does the AI model hold onto data? When’s it deleting it? Where does it store it? You could get into a data sovereignty discussion depending on how far you go down here.

The other important thing is the security standards of the AI model. But if you’re interacting with one of the big AI companies, firstly, they’re not going to change their security model for you, so it’s about you being comfortable with their security model, where they’re at, and what they publicly state. But look for certifications, for example, ISO 27001 or SOC 2 Type certifications. Make sure the companies you’re dealing with have the necessary security certifications as well.

Leland McFarland: Okay. All right, so let’s say a business owner has discovered an employee has gone rogue, and that employee has already uploaded confidential company documents, customer information, or financial records into an unapproved AI application. What should the owner do immediately, and how can they determine whether the information has been compromised?

Tony Anscombe: That’s a great question, because I think that’s going to happen, isn’t it? Somebody’s going to upload something somewhere. The question is whether they admit to what they’ve done. It’s like somebody that clicks the phishing link: do they want to be the person that goes to their IT team and admits they clicked it? And the answer is that’s challenge number one.

But if an employee comes and admits that they’ve had this option, firstly, secure the account. Delete the chat history. Take a copy of the history first, then delete it. So delete any data that the platform may have as well. If it’s still an open session, ask the AI to delete the data, and submit a formal deletion request to the platform.

Say, for example, somebody has uploaded a spreadsheet with customer data in it or intellectual property from the company: you can make a deletion request to the AI platform. Now, I’m not sure how quickly they would actually act on the deletion request—I’d like to think quite quickly—but you should have the right for deletion, especially if it contains personal, sensitive information. So make that deletion request as well.

Leland McFarland: Okay. So AI skills can help small teams automate tasks, work more efficiently. We’re now seeing AI tools that can connect to business applications, interact with files, perform increasingly complicated workflows. What should small businesses consider before putting these capabilities to work?

Tony Anscombe: Well, firstly, let’s explain what an AI skill is, because in small businesses, not everybody might understand what an AI skill is.

If you have an AI agent, firstly, think of your AI agent as an employee. Treat them as such. You’re going to ask them to do a new task today, and that task is to help—let’s use an example—maybe do some website coding. But your agent doesn’t know specifically that type of code. So what you do is you download an AI skill, you attach it to the agent, and suddenly the agent has become really clever and it can do this piece of website coding, probably faster than you could train an employee to do it.

Now, there are lots of these skills around. I used website coding and development as the example because that is primarily what a lot of skills are being used for—more on the technology side. But there are some great skills out there, Leland. For example, you can teach an agent to create PowerPoint presentations for you: you upload a document, it will create the deck, and it will create the speaker notes. So it kind of does the whole package. Employees might start using these types of things.

Now, to give you an indication of where this is going, some AI skills then have a malicious intent. Some of them might not be intending to be malicious; they might be accidentally malicious. For example, a red team may have created a skill to do credential theft. You load that up onto a public portal, and somebody downloads it; it could be misused and it becomes malicious.

However, we’ve also seen malicious skills out there as well that self-obfuscate or download other code. And suddenly you’ve got these bits of code running in your organization that came from a download, and now it’s downloading other code from external sources. You don’t need to be a genius to work out that you’re running unverified, untrusted, unapproved code in your organization. Now, that’s got to be a bad thing.

So when companies look at the AI tools that employees or staff can use, they need to make sure that there’s a verification process—a process of making sure that the code is good and that the tools are approved. Otherwise, you’re likely to get into trouble—maybe not today, but maybe tomorrow—when that code self-obfuscates or self-modifies or downloads something you didn’t intend, then it could be a real problem.

And now, just out of interest, we scanned publicly listed AI skills. We scanned about 900,000 of them over a four-month period, and we found about 3,000 malicious ones. That doesn’t sound like a big number, but you don’t want to be on the wrong end of that 0.3% because it could be nasty.

Leland McFarland: Yeah, any risk is a risk. And it’s only going to go up when more people realize, “Oh, I could really cause some havoc or maybe even make some money by injecting my own little malicious intent in this skill.”

Tony Anscombe: Mhm.

Leland McFarland: All right. So as AI agents become more autonomous, they can go beyond generating content and potentially take actions, such as sending emails, updating customer records, and accessing financial systems. How should small businesses decide what permissions to give these agents, and which actions should always require human approval?

Tony Anscombe: Well, firstly, I said treat the agent like it’s an employee, and I think that’s primarily where you start this conversation. When somebody joins an organization, you only give them access to what they need. Somebody that joins the organization in finance might not get access to the HR database, and the HR person doesn’t get access to the finance details. You already silo people; AI agents are no different. You want to silo what they have access to in the same way that you would an employee. Think of the agent as an employee and say, “This is its task. This is what I want it to do, and what will it need access to?” Once you’ve defined that access model, then you’re already some way to restricting what it can do.

Now, it can read the data in those areas; it might be able to write drafts and stuff like that. But then when do you need to think about making an agent come back to you and ask for human approval?

Firstly, this is something I would never do, but just say you allowed an agent to make a financial transaction, or move funds from one account to another, or adjust price lists, or adjust documents that are commercially sensitive (i.e., that might change what you’re publishing about products). To me, that requires human approval. Personally, I would stick to that being a human task, not just human approval. But companies may automate some of this, and they may go down this route.

Also, agents shouldn’t be allowed to do bulk modifications. When I say that, for example, if you’ve got a customer database, you don’t want to suddenly find the customer database has only got half the data in it because the agent felt that half of it was irrelevant, so the agent deleted it. Keep the agent within the tasks specified. Don’t allow it to do things that you wouldn’t allow it to do sensitively.

In fact, here’s a great question: somebody asked me this yesterday, Leland: would you allow an AI agent to make modifications to your own bank account?

Leland McFarland: Ooh, no, no, no, no!

Tony Anscombe: You’re shaking your head, I’m shaking my head. Neither of us would. Don’t do it at work either. Only allow AI to do at work what you’d allow it to do in your personal life.

Leland McFarland: All right. So, one finding from ESET’s research that stood out to me was that 87% of U.S. businesses surveyed felt confident in their cybersecurity resilience, yet 54% experienced at least one cyber incident during the previous year. Why do you think there’s such a disconnect between confidence and actual experience, and how can business owners determine whether their confidence is actually justified?

Tony Anscombe: Well, firstly, I think the two things are related. Now, let’s talk about the incident first. If you have an incident, it depends how we define what an incident is. If an incident is a mass ransomware attack, they probably don’t feel very confident about actually being business resilient from it. If the incident is somebody clicking on a phishing link, then the business may well feel confident of how to recover from it. So I think there’s an element of understanding how somebody defined what an incident is in their mind.

Also, companies now expect an incident. I think it’s become normal for a company to understand that we will have incidents through the year. The issue is actually making sure that we don’t have a catastrophic or severe incident. So therefore, their ability to recover is pretty good. If somebody clicks a phishing link, there’s probably some procedure; the IT guy is pretty aware of it, even if it’s what I define as a generalist IT guy—a very small business with one person doing all tasks—who probably knows how to deal with it, and the business will recover.

So I’d probably say the confidence matches the expectation that there will be an incident, but it not being a catastrophic incident.

Leland McFarland: So another interesting finding was that 29% of U.S. respondents identified complexity and integration challenges as their biggest barrier to improving cybersecurity, compared to 19% who cited budget limitation. What does this tell us about the challenges businesses face when trying to improve their security, and what can they do to make cybersecurity more manageable?

Tony Anscombe: Well, I think for small businesses—because we are talking primarily about small businesses here—modern security architecture and looking for threats in businesses has changed significantly. Realistically, when an incident unfolds beyond what I’d say is something trivial, you need a security analyst on the other end.

So I would recommend for small businesses looking to protect endpoints who take this seriously: unless you have security analysts on hand (i.e., you’re big enough to have a team that includes specialized people), I’d actually outsource this and run MDR—I’d run a managed service. And the reason being is because what you want is an analyst sitting there at 3:00 a.m. on a Sunday morning when a cybercriminal unleashes something bad. You want an analyst that understands, “No, that’s not good,” and actually a system that will automatically block but raises an alert that might cause the analyst to isolate.

Could you do that internally in a small business? I think that’s challenging in today’s environment, and I’d strongly recommend any small business go down the MDR route.

Leland McFarland: Yeah, I’ve been that guy at 3:00 in the morning fighting off nefarious fights. It’s not fun. It isn’t.

All right, so we’ve covered a lot of ground today, particularly when it comes to AI and new security challenges it creates. If a small business owner does just one thing after watching this interview to make AI safer, what should it be?

Tony Anscombe: Well, I’m going to give you two things, Leland, because I can never just think of one on its own.

One is you need to talk to a company that’s actually securing AI. For example, ESET: we have a lot of technology in there that secures interactions with AI, AI conversations, AI agents, etc.

But if somebody has done nothing, and they’re listening to this and go, “We need to do something,” the first thing I’d recommend you go off and do actually is audit what is happening in your organization. Who’s using what, how it’s being used, and understand what the pain point is that you then need to try and address. I think knowledge of what’s going on is one of the first places to start.

Leland McFarland: All right. Well, that is what I have for you today. Thank you so much for coming on and sharing all your knowledge. Where can they learn a little bit more about ESET and what they can offer for business owners?

Tony Anscombe: Well, we are a cybersecurity company; we’ve been around for 30 years. Our website is eset.com, so nice and simple. We have everything, actually, all the way from consumer products all the way up to enterprise managed services or EDR, and we do threat intelligence, etc.

But for that small business section, I’d look at our small business products, and I’d also look at our MDR offerings, and I’d focus on that middle tier of what we offer in there.

Please, as a small business, stay away from our consumer offerings. They’re designed for consumers, and I know some small businesses—very small businesses—will go down that path. There’s a reason business products are business products: because they come with management, deployment, etc.

And they also come as packaged solutions. For example, one of the big things today is vulnerability patching. We’ve seen some big vulnerabilities out there recently—PHP and other things. So vulnerability and patch management, encryption, MFA—it comes as a package, and I think that’s super important that somebody looks at it as a package.

Leland McFarland: All right, well, thank you for coming on. I appreciate it. And yeah, until next time.

Tony Anscombe: Look forward to it. Thank you, Leland.

My conversation with Tony Anscombe highlighted an important reality for small business owners: adopting artificial intelligence is no longer just about finding ways to improve efficiency. It also requires understanding the security risks that come with allowing AI tools to access company information and perform business tasks.

One of the biggest takeaways from our discussion was that businesses don’t necessarily need complicated security systems or extensive IT resources to begin addressing AI-related risks. As Tony explained, something as straightforward as a one-page AI usage policy can help employees understand which tools are approved, what information should never be shared, and why AI-generated content needs to be verified before being used.

His warnings about AI agents and downloadable skills were particularly noteworthy. With ESET’s research identifying approximately 3,000 malicious AI skills among roughly 900,000 examined, the importance of verifying third-party tools becomes clear. Tony also emphasized that AI agents should be treated much like employees, receiving access only to the systems and information necessary to perform their assigned tasks.

Another important point was that small businesses should not confuse confidence in their cybersecurity with complete protection from threats. As cyberattacks become more sophisticated, having the right expertise available to identify and respond to incidents can be just as important as having preventative security measures in place.

Perhaps the most practical advice came near the end of our conversation, when I asked Tony what small business owners should do first to make their use of AI safer. His recommendation was to start with an internal audit to understand which AI tools employees are using, how they’re using them, and what risks those activities may introduce.

For business owners who have already adopted AI or are considering expanding its use, that is a manageable first step. Understanding where AI is being used makes it easier to establish safeguards, choose appropriate tools, and make informed decisions about how much responsibility to give automated systems.

Small businesses can learn more about cybersecurity and AI security practices through ESET, its WeLiveSecurity publication, and the National Institute of Standards and Technology’s AI Risk Management Framework.

As AI capabilities continue to evolve, businesses that make security part of their adoption strategy will be better positioned to take advantage of the technology while protecting the information, operations, and customer relationships they depend on.

This article, “ESET’s Tony Anscombe on Shadow AI and Cybersecurity Risks for Small Businesses” was first published on Small Business Trends

SHARE IT: