At Computex 2026 in Taipei, the agentic AI track became the stage for a pragmatic advance: on-device deepfake detection that never leaves your PC. Inside Qualcomm’s booth, San Francisco-based Scam.ai unveiled Halo, a live video-analysis model designed to flag synthetic faces during video calls, shifting anti-fraud protection directly to the endpoint.

The Qualcomm tie-up and local processing

More than a branding exercise, the partnership gives Scam.ai access to Qualcomm’s device ecosystem and optimization tooling, enabling Halo to run on personal computers without relying on cloud infrastructure. The model operates unobtrusively in the background during any video conferencing session, analyzing the feed in real time and alerting if a face appears AI-generated or tampered with. Crucially, no video data leaves the user’s machine.

This matters acutely for HR and recruiting teams, where identity fraud in video interviews has surged by over 2,000% in three years, yet only 31% of HR leaders feel equipped to spot it. It also matters for CEOs, CFOs, and venture capitalists who routinely take high-stakes calls where a deepfake could trigger financial or reputational damage.

Why local inference reshapes the security model

Traditional deepfake detection leans on cloud pipelines: video frames are uploaded, analyzed on server GPUs, and results sent back. That path introduces latency, bandwidth costs, and – most critically – a privacy blind spot. In enterprise settings where calls may involve personal data or trade secrets, sending a video stream to a third party is often a non-starter under GDPR or internal compliance rules.

Halo flips the architecture. Inference runs on-device, leveraging Qualcomm’s AI-optimized processors, and remains entirely passive – no changes to the user’s workflow, no extra steps. The result is a transparent security layer that screens for synthetic video without compromising confidentiality or adding friction.

The bigger picture: data sovereignty and the personal AI shift

The announcement aligns with a broader industry current: AI workloads migrating from hyperscale clouds to personal devices. Qualcomm’s Snapdragon X Elite platform and the Copilot+ PC initiative are pushing local AI from productivity to security. On-device deepfake detection is not a standalone gimmick; it’s a building block of an architecture that prizes data sovereignty and reduces cloud dependency.

For organizations weighing on-premise or edge deployments, an optimized model tied to a specific hardware ecosystem raises valid questions about portability, minimum system requirements, and total cost of ownership compared to a cloud API service. While Scam.ai hasn’t released detailed technical specifications, the co-engineering with Qualcomm suggests a path toward simpler large-scale adoption – one that keeps data firmly under the organization’s control.

What’s next

Halo will be available starting June 2026, with enterprise integration details and additional platform partnerships expected in the months that follow. The signal is clear: fighting deepfakes won’t rely solely on user awareness, but on near-invisible, privacy-first technologies embedded where conversations happen. As video fraud attempts multiply, tools like Halo could quietly redefine what it means to trust a face on screen.