Managing AI’s security risks

What are the risks of using AI?

The most well-known risks are data leaks and cybersecurity threats. “But for small businesses, it’s more about how people are using the AI,” says Llewellyn. 

Problems can arise from careless use, security weaknesses in the systems involved, or even criminals using AI to target your business. 

The key is awareness, says Llewellyn.

“The goal is not to eliminate all risk. It's about understanding your planned use and implementing appropriate safeguards.”

Sensitive information can leave your control

One of the most immediate concerns is also one of the simplest: an employee sharing sensitive information with an AI tool that your business has not vetted or approved. For example, they may upload a document without realizing that it contains confidential customer information or intellectual property (IP).

This kind of unapproved use is often called “shadow AI,” says Llewellyn.

The employee may simply be trying to work more efficiently or use a familiar tool. But if the tool has not been vetted, you may not know what information the employee is sharing or how the provider can use it.

If an AI provider uses your information to train its models, sensitive details could become part of what the model learns. In some circumstances, those details could then appear in responses to other users.

Similarly, information stored by the provider could also be exposed through a security breach or unauthorized access. The longer the provider keeps it, the greater the risk. 

Scams and impersonation

According to the Cyber Centre, criminals can use AI to create convincing messages or imitate someone’s voice or appearance. A request that seems to come from a colleague or executive may be an attempt to steal money or confidential information.

Manipulation of AI tools

Attackers can also try to take control of AI tools. For example, they may hide malicious instructions in content that an AI tool (like a chatbot) reads. This is known as prompt injection and can cause the tool to behave in unintended ways.

Inaccurate output and overreliance

AI can produce plausible but invented information, commonly known as hallucinations. These are not necessarily security breaches, but relying on incorrect output is still a risk that can harm your business.

This problem can be amplified if a tool is able to act on its faulty output without human oversight.

How to protect your business from AI risks

Fortunately, there are ways to minimize all these types of risk.

Understand what your AI will access and do

Start by examining your business processes, says Llewellyn. Break down each process, such as responding to customer inquiries, into individual tasks and then identify which tasks you want AI to handle. For each one, before choosing a tool or connecting it to your systems, make sure you understand:

• What information the AI needs to complete the task
• How the tool will use that information
• Where the information and AI-generated results will go
• Whether the tool will only make suggestions or take action
• Who will review the tool’s work

For example, using a generative AI tool to brainstorm with nonsensitive information carries much less risk than using it to process confidential financial or employee information.

Use the level of risk to determine the safeguards you need. 

Avoid giving an AI tool more access than necessary, says Llewellyn.

Even at a very small company, it’s ideal to have an employee or consultant who knows what to look for and what to avoid, he says.

“Ideally, there should be someone who knows to ask questions like, ‘Are we using encryption? Which settings need to be turned on or off? Are the correct logs coming in?’”

Check the tool before sharing information

Before choosing an AI tool, read the provider’s terms of service and its security and privacy documentation, recommends Llewellyn. Review the data retention policy to understand what information it keeps and for how long. Also check whether the provider can use your data to train its AI models. 

Don’t assume a paid AI subscription is more secure than a free one. Business or enterprise plans may offer stronger security controls and privacy protections, but you still need to check the plan’s security features, privacy settings and terms of service, says Llewellyn. 

“At the end of the day, you’ve got to read the fine print.”

Key questions to ask before adopting a new AI tool

Review the tool’s settings as well as the contract. Make sure the tool’s configuration matches your intended use.

Apply these checks to both new tools and AI features built into software you already own. Established business systems may have undergone extensive security testing, but you should still assess whether their AI features meet your needs.

Knowing and trusting a product is not a substitute for understanding its AI features.

Give employees clear rules and oversight

Security also depends on how employees use the tools you’ve approved.

“Staying current with what’s going on in AI is like a full-time job,” says Llewellyn. “But all employees of any organization should have some level of AI knowledge.”

He recommends training employees on what they should and should not do with AI. 

Training should establish clear rules about:

• which tools and uses are approved
• what types of information employees may share
• who to ask when they’re unsure
• which outputs and actions require human approval

The Cyber Centre recommends that businesses create an internal policy for acceptable AI use.

Match the training you offer to people’s responsibilities. Most employees need practical instructions, but those responsible for security should have a deeper understanding of permissions, settings and monitoring.

Llewellyn also recommends limiting automation when you’re just getting started.

“You really want that human in the loop,” he says.

Keep basic security protections in place

AI security builds on the protections your business should already have.

Llewellyn encourages businesses to concentrate on cybersecurity fundamentals, like updating software regularly and using multi-factor authentication.

The Cyber Centre’s IT security guidance also emphasizes giving employees only the access they need to do their jobs. 

Make verification a habit. If someone asks for an unusual payment or confidential information, always confirm the request using contact details you already trust. 

Start small and match safeguards to the risk

When you first begin using AI, start with a clearly defined task that involves nonsensitive information. Set rules, train the employees involved and assess the tool’s performance before giving it more responsibilities.

Use stronger controls for tasks involving sensitive information. Keep track of what a tool can access, require human approval where needed, and review critical steps before the tool takes action.

The goal is to use AI with confidence while keeping employees accountable and business information protected.

“AI is a business tool. It's a security tool. It's an everything tool,” says Llewellyn. “Don’t shy away from it, but do make sure you are cyber secure and have people who know what they’re doing.”

Next step

Discover how to reduce your risk of cyberattacks by downloading BDC’s free guide Addressing Cybersecurity in Your Business.