SafeToNet x Intel

Safeguarding, optimised for the AI PC

SafeToNet and Intel are collaborating to bring private, real-time and preventative safeguarding to PCs at global scale. Combining SafeToNet’s pioneering child-safeguarding technology with Intel’s platform engineering expertise and one of the world’s most established computing ecosystems, the collaboration gives PC manufacturers a practical route to building protection into computers used at home and in classrooms around the world.

HarmBlock helps prevent children from seeing, filming, storing or sharing harmful sexual content, including child sexual abuse material. Built-in safeguarding can protect children across screens, cameras and stored visual content, rather than being added later as another application families must find, install and manage.

Through the collaboration, SafeToNet has optimised HarmBlock for Intel-powered AI PCs. It runs efficiently on the NPU in compatible Intel® Core™ processors, providing manufacturers with a proven technical foundation for integrating private, real-time safeguarding into their PCs from the start.

Discuss an OEM Integration Read the collaboration announcement

Collaboration

The AI PC needs a safeguarding layer

AI PCs are opening up new ways for people to create, communicate and experience digital content. Greater capability should come with stronger protection for children.

Real-time visual safeguarding has to work privately and efficiently in the background, without getting in the way of the everyday user experience. That’s what SafeToNet and Intel have set out to make possible. A collaboration like this reflects how far HarmBlock’s on-device AI has come, and how seriously child safety is now taken.

Safeguarding expertise

SafeToNet brings HarmBlock, its on-device visual AI, and the experience needed to bring it into real-word devices.

Platform expertise

Intel brings client-platform engineering, AI acceleration technology and the OpenVINO deployment ecosystem.

A route to manufacturers

By optimising HarmBlock for supported Intel platforms, the collaboration gives PC manufacturers greater confidence and a more efficient route to integration.

Optimisation

More than moving an AI model onto a PC

HarmBlock Optimised for Intel involves work across the AI model optimisation, runtime and continuous safeguarding pipeline.

01 Model optimisation

Quantisation
Smaller model footprint
Reduced inference cost
Safeguarding-performance retention

02 Runtime integration

Intel OpenVINO execution path
CPU and GPU support
NPU validation where supported
Stable fallback between supported resources

03 Continuous operation

Background inference scheduling
Threading and pipeline optimisation
Long-running memory management
Power-consumption optimisation
Stability and stress testing

Screen

Helps prevent harmful sexual content from being seen

A screen integration enables HarmBlock to recognise harmful sexual content appearing on screen. If detected, the content can then be blocked completely or selectively masked, helping prevent it from being seen.

Camera

Helps prevent harmful sexual content from being photographed, filmed or livestreamed

A camera integration enables the device to detect harmful sexual content and replace it with a blocking frame, helping prevent it from being photographed, filmed or livestreamed.

File System

Helps prevent harmful sexual images or videos from becoming accessible on the device

A File System integration enables HarmBlock to recognise harmful sexual content in incoming or stored visual files. Detected content can then be blocked, quarantined or managed in the way most appropriate for the deployment.

Available capabilities and the form of intervention depend on the final device and partner implementation.

Children can encounter harmful content while browsing, receiving files or using a camera, not only through specific websites or applications.

When integrated into a PC, HarmBlock can recognise this content directly on the device and enable it to respond in real time. For example, content can be blocked on screen, prevented from being filmed or livestreamed, or quarantined before it becomes accessible.

For families, this creates the potential for safeguarding that works automatically in the background, without parents having to monitor every digital interaction, and without sending a child’s private content to the cloud for analysis.

How Intel hardware is used

Designed around the resources available on the device.

HarmBlock’s OpenVINO-based execution path is designed to use supported Intel compute resources efficiently, with fallback where a particular form of acceleration is unavailable.

The aim is not to use every processor for every decision. It is to select supported execution paths that provide reliable, efficient safeguarding across different Intel hardware configurations.

The optimisation for Intel's NPU cut inference time by up to 55%, compared with the CPU alone *

SafeToNet’s optimisation work focused on enabling HarmBlock to run efficiently on the NPU using Intel's OpenVINO™ execution pathway.

In internal testing, the Intel-optimised model running on NPU recorded average inference times of 20 to 25 milliseconds per image. Compared with average inference times of 35 to 45 milliseconds recorded during CPU-based reference testing, this represents a reduction of up to 55%.

Up to 55% reduction

NPU • OpenVINO CPU-Based

Run-average Inference Time

*How we tested: we ran the AI model over a set of 100 images, repeated 5 times, on both NPU and CPU, and measured the average time it took to check a single image. Repeating the run gives a reliable average rather than a one-off result. Average times per run were 20 to 25 ms on the NPU and 35 to 45 ms on the CPU. The “up to 55%” figure compares the fastest OpenVINO average with the slowest CPU average, and is rounded down. Real-world performance varies with the model version, hardware and software configuration, and workload.

Privacy

Private by design

HarmBlock is designed to preserve a child’s privacy while it protects them. It is there to recognise harmful sexual content and not to identify the child, follow their activity or create a record of what they see. Safeguarding decisions are made directly on the device, so visual content does not need to be sent elsewhere for analysis.

Intel platform optimisation changes how HarmBlock uses device resources. It does not change HarmBlock’s core privacy architecture.

HarmBlock on device

Operates without a network connection
Does not require remote content analysis
Does not identify the child
Does not collect or store user data
Does not save nor create a library of analysed content

For OEMs

A practical foundation for OEM implementation.

The main goal of the SafeToNet and Intel collaboration is to make HarmBlock integration straightforward for the manufacturers we work with.

Differentiate the device

Offer a meaningful safety-by-design capability for family and education products.

Reduce foundational engineering work

Begin with an Intel-optimised model and runtime path rather than having to explore optimisation methods from scratch.

Adapt the experience

Choose relevant HarmBlock modules and determine how the final device communicates and responds to safeguarding interventions.

Preserve privacy

Enable visual safeguarding without sending sensitive content off the device.

Support different hardware configurations

Use supported execution paths and fallback appropriate to the selected Intel platform.

Create a benefit families can understand

Offer safeguarding that can work across screens, cameras and stored content without requiring families to find, install and manage another application.