GenAI data loss prevention
We stop sensitive data from reaching AI tools. The protection builds on the general data loss prevention used in IT security and applies specifically to traffic going to AI services.
The risk
What goes wrong when data ends up in an AI service
When an employee pastes a draft contract or client data into a public AI service, that data leaves the organization. We covered the other AI risks in our article Security risks of enterprise AI adoption.
Personal data
Under the GDPR this is a data transfer, and it typically happens without a legal basis or a data processing agreement.
Trade secrets
The information can lose its protection if it has been uploaded to a public service.
Supply chain
The NIS2 supply chain measure also applies to the AI provider if the service has become part of the organization’s processes.
Shadow AI
A department pays for an AI tool subscription by card, uploads its own data, and nobody knows about it. IT cannot write a data loss prevention rule for something it does not know about.
Rollout
How we roll it out
-
Data that must not go to AI
We define which categories of data must not enter an AI service, such as client data or draft contracts. If a data classification already exists from general data loss prevention, we reuse it.
-
Rules for AI services
We set up rules for the prohibited data categories on traffic going to AI services.
-
Starting in monitor mode
In this phase the rules only log. The log shows what data would be sent to AI and which rule would get in the way of a legitimate workflow.
-
Blocking
The tuned rules then stop prohibited data before it reaches the AI service.
-
Report
We produce a report on the prevented cases. Operation after the rollout is available under a custom agreement.
Deliverables
What you get at the end
items handed over
- Definition of the data that must not go to AI
- Data loss prevention rules for AI services
- Start in monitor mode, then blocking
- Report on the prevented cases
Related services
What goes with it
Traffic to AI is only one channel among several, and using AI involves rules that go beyond data.
Data loss prevention (DLP)
Protection for sensitive data on the traditional channels.
AI security controls deployment
Rules on which AI services are allowed and which are prohibited, plus filtering of malicious instructions.
Secure Copilot and Microsoft 365 AI rollout
Copilot sees only what the user can see. Before the rollout we put permissions, sensitivity labels, and data protection rules in order.
Frequently asked questions
The questions we hear most
Why do we need separate protection for AI if we already have data loss prevention?
General data loss prevention monitors email, cloud storage, printers, and removable media. Those rules do not cover traffic going to AI services, so that traffic needs rules of its own.
Wouldn’t it be simpler to block AI services?
A ban on its own does not work, because employees find ways around it. It needs a usable alternative alongside it: an enterprise subscription with a data processing agreement, or a locally run model that keeps the data in-house. You also need to make clear what may be entered where.
How do we find out which AI tools employees use?
Proxy or firewall logs can show which AI services are reached from the organization, and in what volume. The full picture, meaning an inventory of the tools and data flows in use with a risk rating for each tool, comes from the Shadow AI assessment. We covered the topic in our article Shadow AI: when employees use AI without approval.
