Managing Insider Risk With Adaptive Controls

An alternative to hard blocks is adapting protective mechanisms based on user behavior. That means: if you act according to company policy, you shouldn’t face very hard restrictions on what you can do with sensitive company information. But if you act irresponsibly, you should face the consequences.

Below I’ll describe an approach on how this can be made possible in a few steps.

Step 1: Discovery

Most organizations have far more sensitive information than they realize (often in places they were not expecting to find it), so our first step is discovery (automatic and manual).

Example: A colleague opens a Word document titled document1.docx and finds it contains negotiation points around an M&A target.

Step 2: Sensitivity Classification

Let’s start step 2 by acknowledging the fact that different types of information require different types of protection. A credit card number on its own is less sensitive than a credit card number alongside an expiration date and CVV, a product brochure is less sensitive than an email containing a termination notice, etc. Sensitivity classification is how we provide context about how much protection an information asset needs. The classification (i.e., label) guides our people and systems in what they are allowed to do with the asset.

Sensitivity classification is not a new concept. Stamping documents as “Highly confidential” to provide context to people for how they should act around information is something people/organizations have done long before we went into the digital age. Stamping a sensitivity label on a digital document does the same thing, i.e., it communicates sensitivity consistently across people, but it takes it one step further as it also communicates it on a system level (i.e., what are systems allowed to do with this document now that it is classified).

Example: A user classifies a document as “Highly confidential” to indicate that this document should not be exposed to just about anyone and any system.

Step 3: Data Loss Prevention (DLP)

When we have the classification context, we can make very intelligent decisions around how to handle risky activities on higher sensitivity assets. If somebody, for example, attempts to paste a paragraph from a document classified as Confidential into a non-approved LLM, we can set up protective controls that allow us to respond to that exfiltration scenario in real time (i.e., block the activity and warn the user).

Example: An AI agent requests information regarding a potential acquisition. The request passes through a gateway which checks whether this information is appropriate to be returned to the agent. The classification context, Highly confidential (NDA required), is processed and the request is denied.

Step 4: Insider Risk Management

The interesting thing about many data exfiltration scenarios is that no single action is necessarily a red flag. People will make mistakes when they work with information (DLP blocks can and should be used with a warning dialog to provide meaning to a user). The actual warning signs often appear when you look at people’s behavioral patterns over time. A large download may be normal. Prompting AI is normal. Accessing sensitive files is normal. But if a user performs all three of these activities within a short time span, it may require an investigation (as it could be a sign of an incident).

When seemingly normal activities start forming a concerning pattern, Insider Risk Management is our detection tool. For our response: we use Adaptive Protection to apply just and appropriate protective mechanisms. This means that if an identity is exhibiting risky behavior (perhaps it has been hijacked), we can systematically ensure that their account gets stricter controls (we simply adapt what they are allowed to do in alignment with how risky their actions have been), while everyone else continues working as they normally do. A heightened risk signal means it’s time for us to open a case and investigate.

Example: A user downloads a SharePoint library with Highly confidential information. They attempt to upload these documents to their personal Google Drive but the activity is blocked by a DLP policy. They then attempt to email it to their personal Gmail, but again, the action is blocked. The system finds these actions suspicious, adapts the DLP to ensure the user can no longer open documents labeled Highly confidential and is forced to re-authenticate with MFA. The system also alerts the right security staff who is able to open a case and investigate the user.

To Summarize

Together, this combination of capabilities ties a neat, flexible and defensible bow around information protection, and makes addressing the following types of challenges possible in real time (i.e., as they happen):

• Large-scale downloads before an employee leaves an organization
• Company IP being uploaded and processed by untrusted LLMs
• Employees sending sensitive company information to private email accounts
• Confidential company IP being copied to USB sticks
• Mass downloads from SharePoint and OneDrive
• Attempts to move information into collaboration platforms outside company control

The most important takeaway from this article should be that protecting the information our organizations rely on becomes much more effective when IT Security has context about how sensitive an information asset is, and how this context allows us to expand upon traditional perimeter-based protection models.