Your Employees Are Already Using AI. You Just Don’t Know Which One.

There’s a conversation happening in security circles right now that sounds a lot like the one we had about shadow IT ten years ago. Back then, the problem was Dropbox. People got tired of emailing files to themselves, so they just started using it. IT had no idea, data was leaving the building, and by the time anyone noticed, it was everywhere.

AI is doing the same thing, faster, and with higher stakes.

The Problem Nobody Wants to Say Out Loud

Here’s what’s actually happening at most mid-size companies right now: an analyst is summarizing a client contract in ChatGPT because it’s faster than doing it manually. A developer is pasting internal API code into an AI assistant to debug it. A salesperson is feeding prospect data into some AI tool they found on Product Hunt to draft outreach emails.

Nobody told them to do this. Nobody told them not to. And nobody in IT or security has any idea it’s happening.

This is shadow AI, which is the use of AI tools inside an organization without the knowledge or approval of anyone who’s supposed to be managing risk. It’s not malicious. It’s just people doing their jobs with whatever works. The problem is that “whatever works” is sending your data to infrastructure you don’t control, under terms of service your legal team has never read, with no audit trail.

What’s Actually at Risk

The instinct is to frame this as a data loss problem, and it is, but it’s more specific than that.

When an employee pastes a customer contract into a free AI tool, that text goes somewhere. It might be used to train a future model. It might be stored on servers with security controls that are worse than yours. It might be subject to a data breach you’ll never hear about. The employee doesn’t know any of this and honestly shouldn’t be expected to.

If you’re in healthcare, that’s a HIPAA exposure. Financial services, FINRA. Government contracting, CUI handling requirements. Any of those scenarios can turn an employee trying to be productive into a six-figure compliance problem.

The IP angle is worth taking seriously too. Source code pasted into AI tools, internal product roadmaps, M&A analysis, proprietary formulas. Once that’s out, it’s out. You can’t un-train a model.

And then there’s the more subtle issue: inconsistency. If fifteen people on your team are using fifteen different AI tools with no standardization, you have fifteen different levels of accuracy, fifteen different failure modes, and no way to know which outputs to trust.

Why Blocking Doesn’t Work

The obvious response is to block it. Block ChatGPT at the firewall, put it in the acceptable use policy, done.

Except people use phones. They use home internet. They use VPNs. You can make it harder, but you can’t make it impossible, and trying to creates a different problem: you’ve now driven the behavior underground without reducing it. People who feel like they’re doing something wrong are less likely to ask questions or flag problems.

The uncomfortable truth is that the demand for these tools is real. They make people meaningfully faster and better at parts of their jobs. If your organization’s position is “no AI, full stop,” the outcome isn’t compliance. It’s a workforce that’s quietly non-compliant and a competitive disadvantage against everyone who figured out how to do this safely.

What Actually Helps

The organizations that are getting ahead of this aren’t trying to eliminate AI use. They’re trying to channel it.

The starting point is an approved tools list with real enterprise agreements, not just a list of names, but actual vendor contracts that include data protection clauses, no-training commitments, and defined data handling requirements. Enterprise agreements for tools like Microsoft Copilot, Claude for Work, or Google Workspace AI exist specifically to give organizations these guarantees. The free tier of any AI product was not built with your compliance requirements in mind.

Alongside that, you need data classification that actually connects to AI usage. The relevant question isn’t “is AI allowed?” It’s “what data is allowed to go where?” Public information, internal communications, confidential client data, and regulated personal information all need different answers. Building that into policy gives employees a framework to make their own decisions rather than a blanket prohibition they’ll quietly ignore.

DLP controls help, but they’re enforcement, not policy. Training is what changes behavior. And the most effective training doesn’t lead with “here’s what not to do.” It leads with why the risk is real, in terms that aren’t abstract. Most employees genuinely don’t know that the free tier of an AI tool might use their inputs for model training. When they understand what’s actually happening to their data, the decision to use the approved tool instead isn’t a burden. It’s obvious.

The Security Assessment Angle

For organizations that want to understand their actual exposure before they start building controls, the question to answer first is: what’s already out there?

That starts with visibility. DNS and proxy logs showing what AI domains employees are hitting, what data classification those requests might involve, and what the gap is between current usage and any existing policy. In a lot of organizations, that exercise alone is clarifying. The problem stops being abstract when you can point to specific tools, specific teams, and specific data types.

From there, you can build policy and controls that reflect what’s actually happening rather than what you assumed was happening.

The Bottom Line

Shadow AI isn’t a rogue employee problem. It’s a governance gap that organizations created by moving too slowly on AI adoption while the tools kept getting better and more accessible. The employees using unauthorized AI aren’t trying to create risk. They’re trying to do their jobs.

The organizations that respond to that with prohibition are going to lose twice: once to the risk that doesn’t go away, and once to the productivity gap that grows every month they wait.

The ones that respond with real governance, approved tools, clear data policy, actual training, and controls that match the risk, end up in a better position than if the problem had never come up. They know what tools their people are using, they have agreements that protect them legally, and they have a workforce that understands why it matters.

That’s a better outcome than hoping nobody pastes your client list into something you’ve never heard of.

Brackish Security helps organizations assess and address AI security risks, including shadow AI exposure, AI system vulnerabilities, and AI governance gaps. If you’re not sure what your team is using, that’s probably worth finding out.