AI/ML Usage and Governance in Netskope Products

Updated July 2026

A number of capabilities in the Netskope Platform are implemented via AI/ML models. Most of the models are internally developed by Netskope, proprietary to Netskope, and targeted to cybersecurity use cases. AI/ML models have been part of the production Netskope One platform since 2019.

Netskope has an internal AI Governance Committee that sets internal policy, processes, and standards for the safe and ethical use of AI/ML within Netskope and in Netskope products. The Committee includes technical, privacy, and legal experts.

Netskope products that depend on AI/ML models include, but are not limited to:

• Advanced Data Loss Protection (DLP)
• Advanced Threat Protection
• Advanced User and Entity Behavior Analytics (UEBA)
• AgentSkope
• AI Gateway
• AI Guardrails
• AI Red Teaming
• Borderless SD-WAN
• Cloud Confidence Index (CCI)
• Device Intelligence
• Digital Experience Management (DEM)
• Enterprise Browser
• Netskope Private Access (NPA)
• Next-Gen Secure Web Gateway (NG-SWG)
• SkopeAI Data Protection
• SkopeAI Threat Protection

Although some aspects of these products depend on AI/ML models, many other aspects do not. On request, Netskope can provide additional details about specific features and their connection to AI/ML models.

Customer Data Use

Customer data is only used in two situations:

• Per-user models (found in Advanced UEBA, Borderless SD-WAN Operations, and Device Intelligence).
• The Train Your Own Classifier in Advanced DLP.

These are explained here.

• Features with per-user models build a dynamic model of each user’s behavior. The model then aids in identifying anomalies in user activities that may indicate security issues. Each such model inherently depends on the specific user’s data, since a model of a generic user is ineffective at identifying genuine anomalies. Although the model is specific to a user, it is built from relatively low-level transactional information (roughly a clickstream) and does not contain any of the content that is being uploaded/downloaded.
  To limit this usage of your data, you can choose to not use any feature that depends on per-user models (Advanced UEBA, Borderless SD-WAN Operations, or Device Intelligence).
• The Netskope Train Your Own Classifier feature allows you to build your own ML model for DLP that is then usable within the Netskope Cloud for data protection. Each such model inherently depends on data provided by you.
  To limit this usage of your data, you can choose to not use Train Your Own Classifier.

Although there are many other AI/ML models underpinning different aspects of Netskope functionality, all of the other models are trained entirely on Netskope proprietary data or data sourced from vendors.

No customer data is used to train Netskope models except for these cases.

Model Training

As noted above, there are only very limited ways in which customer data ever affects AI/ML models. None of the models based on your data are ever shared with other customers or used in other tenants; they remain exclusively with the original tenant.

At Netskope, we use proprietary data to train and test our machine learning (AI/ML) models, which cater to various applications such as DLP, Threat Protection, UEBA, and NG-SWG. Our model algorithms encompass a wide range, from deep learning neural networks to decision tree-based learning methods. We ensure the accuracy of our models through rigorous lab tests and evaluations in production environments.

Key metrics for assessing model accuracy include precision, detection rate, and false positive rate. Additionally, we take into account model runtime and latency when selecting models for deployment. Our priority is to choose models that offer the highest accuracy and lowest latency for implementation in production.

When a model is released into production, we proactively monitor its performance using both production data and testing samples. Ground truth labels are gathered from various sources, such as:

• Online services and threat feeds, like VirusTotal for malware, and PhishTank for phishing URLs.
• Manually reviewed samples from Netskope’s own corporate network traffic.
• Customer feedback (with their permission).

In the event that we observe a decline in model efficacy, we collect more training data and retrain the model, repeating the training-validation-deployment-monitoring process. Our models are typically updated every six months, or as needed (like after a significant increase of false positives, or a drop of detection rate is noticed). Information regarding model updates can be found in the Netskope Security Cloud Platform services release notes. For example, we update the data protection and threat protection models multiple times a year to keep them up-to-date.

Netskope works to include diverse training data from different sources. However, our training data may not represent 100% of the data the models see in production. For example, we don’t have passport images from every country in the world to train our DLP image classification model. In addition, there may be noise in the training data. For example, a Windows executable file labeled as benign today may turn out to be malicious next month. As a result, the model may not have 100% coverage and accuracy.

Ongoing Netskope inventions may reduce or even eliminate some of these training limitations, and our goal is to avoid any impact on our target security use cases.

Ensuring AI/ML Models are Unbiased

Netskope’s models are essentially agnostic to user characteristics except in the narrow sense of their network traffic. Netskope has developed custom proprietary ML models for use only within the Netskope Security Cloud Platform services to implement the security functions of our platform. These models are related to security features and capabilities for traffic between end-users and web/cloud services. Although users may be distinguishable and recognizable in some models, the models don’t use customers’ personal/PII data, and don’t involve age, gender, race, or other legally-protected characteristics in their security-related decisions. There are no mechanisms inside the Netskope Security Cloud that attempt to determine any user’s legally-protected characteristics and have no controls that relate to such capabilities.

Alignment with Risk Management Frameworks

Netskope utilizes risk management frameworks including NIST AI-100-1 to ensure the design, development, and deployment of ML models is aligned to common framework standards to ensure the many risks of AI/ML are appropriately assessed and managed from a security, privacy, ethics, and responsibility perspective. We are in the process of aligning with ISO 42001, but we are not yet certified compliant.

All Netskope ML models are embedded in Netskope services. Netskope is responsible for both the development and appropriate usage of the ML models, as well as testing for efficacy and replacing models when needed.

In particular, neither customer admins nor end-users can determine whether a given feature or behavior is implemented by an AI/ML model, or whether that model has changed, in the absence of information from Netskope. Accordingly, most of the risk management framework for AI/ML models translates into risk management of the Netskope One cybersecurity platform.

Combining Automation with AI/ML Models and Human-directed Policies for Efficient Decision Making

Netskope’s AI/ML models are integrated into various products, like SkopeAI for data protection and threat protection, Advanced UEBA (User and Entity Behavior Analytics), Borderless SD-WAN, and Device Intelligence. These models are designed to automate the analysis of vast amounts of data to identify anomalies and potential threats by sophisticated scoring and categorization techniques.

While these models can automatically score and categorize data for specific tasks, the decisions regarding their deployment, the configuration of security policies that leverage the model outputs, and their governance are directed by human insight and expertise.

Use of LLMs

The Netskope One platform uses a Netskope-proprietary LLMs for the following use case:

• To inspect prompts and responses in generative AI (GenAI) traffic when users or agents are interacting with GenAI applications or agents. This function is part of both the AI Security Guardrails and AI Gateway products.

The Netskope One platform also uses some third-party LLMs. In every such case, no customer data is logged by the third-party provider, and the third-party provider does not use Netskope customer data for training their models. Third-party LLMs support the following use cases:

• To enhance alert insights in the Device Intelligence product. Only public alert signature names are shared with the LLM, ensuring no customer data is involved. 
• To support conversational interactions with Netskope’s proprietary CCI database, via the CCI Insights Agent of the Cloud Confidence Index product. 
• To support conversational interactions with Netskope’s database of telemetry and metrics, via the DEM Data Intelligence Agent of the Digital Experience Management product.
• To perform end-to-end data security analyst workflows of investigation and remediation, via the DLP AISecOps Agent.
• To perform triage and analysis of insider threats, via the Insider Threat AISecOps Agent.

Note that LLMs and agents are different. Some LLMs are used in features that are not agents, while some agents are not implemented using LLMs. In particular, the following Netskope agents do not use LLMs currently:

• Private Access AIOps Agent.
• DEM Data Intelligence Agent.

Approach for Future Innovations with AI/ML Technologies

We continually experiment with cutting-edge AI and ML technologies, including generative AI and LLMs. New applications of these technologies will be announced as they become available.

Our approach to features that may involve customer data follows best practices, like opt-out by default, clearly indicate the use of customer data, and provide adequate advance notice.

Netskope is committed to maintaining transparency and upholding our customers’ privacy and security.