When I was growing up, my Mom thought I played video games too much (she was right). She didn't want these games getting in the way of my school, and wanted to make sure I still did other activities like playing with friends and reading books. So when she didn't want me playing them, she'd lock up all the games in a large armoire and hide the key.
I didn't let that stop me. Forcefully locking them away seemed unfair and only made me great at key'n'go seek and "alternative" opening methods. And the worst part? I’d put everything right back where it started so I’d get quite a few sessions in before getting caught.
The same thing is happening in workplaces today with AI. Adults are just big kids, and the game has stakes now. When someone’s livelihood is on the line, the tool that helps them hit their metrics is going to be harder to lock away, and some of them are going to find the key anyway and not tell you about it. So what do you, simply not ban unapproved tools, let employees continue to use them even though they don’t meet your security requirements?
Yes.
Yes, using these tools can be a security risk. Unequivocally so if they have access to first-party data. But forcefully taking something valuable away is like ripping away a child’s favorite toy.
That’s not to say you give them free rein to paste in confidential company data and patient records into a free Grok account, just like Mama didn’t buy me Grand Theft Auto at 10 years old and give me unlimited playtime. The approach that eventually worked for young me was “Gaming is okay on nights and weekends after homework is done for a reasonable amount of time”, and similar reasonable guardrails can work in the workplace too.
You just need the right tool for the job.
The problems with fully banning tools
It might seem like the best and simplest response to controlling Shadow AI is to squash every AI tool not on your approved list and stop any other ones from getting added. After all, if something like Copilot is already included in the Microsoft subscription and has data controls in place, why not just lock down the one platform and call it good?
Because it’s never that simple.
AI is here, and at least in the short term, adoption is increasing and more tools are starting to appear. And some of them are more than a simple ChatGPT wrapper with prebuilt project settings that can perform exceptionally well in a limited use case. An outright, hardline ban on any tools except a sliver of approved options risks the forbidden fruit effect and users still using them, just being sneaky about it instead.
The forbidden fruit effect
While it may sound counterintuitive, banning something will make it more appealing for some people, and make something suddenly interesting. This forbidden fruit effect is a real risk for banning AI tools, even when the intention is to do the opposite.
We’ve all done this at some point too. Maybe not passing data to a black box tool, but finding the key to the armoire, sneaking a peek into an off-limits door, eating someone else’s labeled food in the fridge… the effect manifests in different ways.
For some people, banning AI tools will drive curiosity, especially if it’s big, bold, and definitive. What makes the tool so dangerous? Maybe it’s overstated? Maybe it’s incredibly useful? And taking a small bite doesn’t hurt, right? Reviewing a social media post that was going to go public isn’t likely to hurt anyone. Maybe a line of code next. Ticket responses with confidential data. That slippery slope can potentially lead to eating the whole fruit and giving the tool access to the entire CRM and codebase.
Obviously this isn’t the norm, and breaches are on a much smaller scale. But the psychological principle still stands: making tools forbidden can lead to curiosity or potentially be a driver for someone to use it. Providing AI awareness training and education can help combat this by explaining the specific risks and reasons for not using a tool or not using specific types of data with a tool.
Sneak a peek
In theory, completely banning unapproved tools should increase security. Bring down the ban hammer and stop it from ever becoming a problem. If it’s not allowed, it’s over, right?
For some employees, yes. They won’t be able to use their work credentials to sign up for or use tools without approval. And that’s the problem.
While you’ll be able to stop some, and maybe even most unapproved AI tools with bans, the survivors get more dangerous. Instead of using locked-down, monitored work accounts, persistent users will switch to private accounts. Or worse.
Here’s a few ideas I came up on the spot that could get around bans:
- Use a personal device.
- Go off network (like a mobile hotspot)
- Copy and paste instead of using a direct integration
- Screenshot data or take a phone photo
- Type everything by hand like a madman
There will be a bunch more than that 30 seconds of brainstorming. Tech-savvy users can come up with much more sophisticated workarounds that can go completely unnoticed.
A real-world ban workaround
I found an interesting ban workaround in a Reddit MSP subreddit. The user was tech-savvy and didn't like the idea of losing all their productivity gains so suggested an idea to their MSP.
The idea boiled down to the user offloading prompts from their device into their home network, having that unmonitored network submit the prompt to the LLM, then have the network send back the prompt results to the work computer. This circumvented the blocking and monitoring the software put in place, and the company would be none the wiser (if the user hadn’t explicitly brought it up with their MSP).
This type of “know enough to be dangerous” can be a huge problem for the organization. In this kind of scenario, the user was competent enough to use their skills and experience to solve their immediate problem. But the solution was also a much bigger risk than sometimes putting bad data into the LLM. Because they weren’t given the right tools for work, they found a way to use them anyway, just without the oversight or anyone in the company knowing what they’re up to.
Unless you have a completely secure, air-gapped state-of-the-art facility that forces extreme security policies, dedicated people who want to use a tool will find a way to use it. And when they do, it will probably be a much more problematic way than using their work computer.
How can you make sure users use AI safely?
Banning something is one thing. Enforcing consequences is another. If two-thirds of senior decision makers are using unapproved AI tools (likely including the majority of the C-suite too), you can’t just fire them or the business (probably) dies. Ignoring the problem leaves the organization no better off, and the risk of sensitive data being exposed to AI only increases with time and use.
So what can you do instead?
My personal recommendation is to allow AI tools for specific use cases by limiting what kinds of data is allowed. Then, combine that with a written, explicit AI usage policy that defines the allowed tools, datasets & content types allowed to be used with AI, and reasons why tools and content types aren't allowed.
Back to the parenting comparison: there’s plenty of stories of frustrated stories of kids not buying the “because I said so” argument. The modern equivalent is using blocking apps without explanation. Sure, you can simply block apps you think are harmful on child devices and lock it down with full monitoring, but if that isn’t accompanied by screen time discussions etc., is it really effective, or does it foster creative ways to get around the bans?
My recommendation: Allow AI tool use with lightweight guardrails
I’ve made our case against outright banning unapproved AI tools, but in the end, it’s up to you handle it. When you see bad behavior are you going to be mad, or just disappointed?
Obviously there are exceptions to this rule. If you’re in a highly-compliant industry and can get fined for having data leave your environment, you’ll have stronger security requirements.
But for most industries, letting your team embrace the AI tools they like and work for them is likely to give the best results. Because what’s the real risk here? That someone says “Take out the swears before I write this to my boss?” No. It’s that sensitive, confidential, or proprietary data is fed to the AI and could be leaked. So stop that part.
Oveka is a great way to make that happen.
Prevent sensitive data from going to AI with Oveka
Oveka was built for MSPs to manage AI monitoring and access across multiple clients from a single interface. Our platform is incredibly simple to set up and lets you define how each tool can be used.
Have a business account with a DPA with OpenAI? Let it be summer vacation with unlimited game time. Want your developers to code their software by hand so they know how it works? Block code from being sent to Claude, but let them ask for advice on how to solve a problem.
You have full control over what data is allowed with AI and what is restricted. Block credentials and Secrets, file uploads, sensitive data, and more, with the option to set different rules for different users or groups. Need to take the hardline and ban everything but the corporate account? You can do that too. You can even allow users to override it so they’re workflows aren’t blocked and have Oveka tell them that sensitive content may be in the prompt and have them confirm it’s okay to send.
To see if Oveka is right for you, get in touch with our team. We’ll give you a demo and set you up with a free trial so you can see how simple it is to find, monitor, and control how data is used with AI. To get started, book a demo now.