Scam.ai launched Halo on August 5, 2026: A desktop app that watches your Zoom, Teams, or Google Meet call in real time and flags a synthetic face before you act on what it told you. It runs entirely on your device, costs $14.99 to $29.99 a month depending on the plan, and is currently Windows only. Here is what it actually does, how to set it up, and where it still falls short.
In January 2024, a finance employee at engineering firm Arup joined a video call with his CFO and several colleagues. He recognized their faces. He recognized their voices. He then authorized 15 wire transfers worth $25.6 million to accounts he was told to use. Every person on that call was an AI-generated deepfake, built from public video and audio of Arup’s real executives.
That case became the reference point for an entire category of fraud, and it is the one Scam.ai leads with in its own marketing. Halo is the company’s answer to it: catch the fake while the call is still happening, not in a report written after the money is gone.
What actually shipped on August 5
Scam.ai, a San Francisco based startup also known as Reality Inc., built Halo in partnership with Qualcomm. The official launch announcement confirms the app is live now for Windows, with other platforms planned. That is a slight shift from an earlier Computex 2026 announcement in Taipei, where Scam.ai had targeted a June 2026 release. The product slipped by about six weeks and landed in early August instead.
Halo sits quietly next to Zoom, Microsoft Teams, or Google Meet. It watches the faces visible on your screen, checks them roughly four times a second, and surfaces an alert if one of them looks synthetic. Scam.ai’s co-founder and CEO, Simiao (Ben) Ren, put the pitch simply in the launch announcement: “We didn’t want to build another tool that tells you about a deepfake after the damage is already done.”
How it actually works
The detail worth understanding here is where the processing happens, because it is the whole point of the product.
Halo runs its detection model directly on the device you are using for the call. According to Scam.ai, video frames are analyzed in memory and discarded. Nothing is recorded, nothing is uploaded, and no video ever reaches Scam.ai’s servers. The app reads the same rendered call window you are already looking at, which is also why it does not need a meeting bot, a browser extension, or a link shared into your call.
Qualcomm’s involvement is about hardware. Scam.ai optimized the detection model to run on the Snapdragon NPU, the dedicated AI chip built into newer Snapdragon laptops, so the analysis happens with essentially no added latency. That is the headline pairing in the press release. What the release does not emphasize, but Scam.ai’s own product page confirms, is that Halo is not locked to Snapdragon devices. It runs on Windows 10 or 11, on both ARM64 and standard x86/x64 hardware, needing only 4GB of RAM and about 1GB of disk space. If you have a Snapdragon-powered PC, detection likely runs faster and lighter. If you do not, the app still works.
Halo currently supports more than 10 meeting apps, including Zoom, Microsoft Teams, Google Meet, WebEx, Skype, Slack huddles, and GoTo Meeting, across both native desktop apps and browser-based calls.
How to set up Halo: step by step
- Check your system. You need Windows 10 or 11, on ARM64 or x64 hardware, with at least 4GB of RAM. There is no macOS or Android version yet.
- Go to the Halo product page and download the app. Scam.ai’s download page gives you the Windows installer directly.
- Install it and leave it running in your system tray. You do not open Halo before a call. It sits in the background and picks up active Zoom, Teams, or Meet windows on its own.
- Join your call as normal. No plugin, no bot joining the meeting, no link to paste. Halo starts checking visible faces the moment it detects an active call window.
- Watch for a live alert, not a rating on every face. Halo deliberately waits for enough evidence before it changes a verdict. A blurry frame, bad lighting, or someone turning their head is not enough on its own to trigger a warning. That is a real tradeoff: fewer false alarms, at the cost of not being instant on frame one.
- Check the local meeting history after the call if you want a record. Verdicts and alerts are saved on your machine, not on Scam.ai’s servers, so you can review who got flagged and when.
- Pick a plan once your trial period ends. Scam.ai lets you download and try Halo before committing to a subscription, then choose monthly, annual, or lifetime access.
What it costs
| Plan | Price | Best for |
|---|---|---|
| Monthly | $29.99/month | Trying it out or a short-term need, like a hiring push |
| Annual | $179.88/year ($14.99/month) | Anyone using it regularly, roughly half the monthly price |
| Lifetime | $399 one-time | Long-term personal use, includes future model updates |
All three plans are purchased inside the app itself. Businesses that want to roll Halo out to a team need to go through Scam.ai’s sales team for volume licensing rather than buying individual seats. Worth noting separately: Scam.ai’s core detection API, used for checking individual images or files rather than live calls, has its own free tier of 200 images a month before usage-based pricing kicks in. That is a different product from Halo, aimed at developers building detection into their own apps, not at someone protecting their own video calls.
Who this is actually for
Halo is built around a narrow, specific decision point: the moment before you act on what a video call told you. Scam.ai’s own positioning names two groups plainly.
Finance and treasury teams. Anyone who can approve a wire transfer, confirm a vendor change, or authorize a payment based on a video call is the exact profile the Arup case describes. If your job involves that kind of approval, running Halo during high-stakes calls costs less than a single fraudulent transfer.
HR and recruiting teams. Interview fraud is a documented and growing problem. Gartner projects that one in four candidate profiles worldwide will be fake by 2028, and remote interviewing has made face-swapped or impersonated candidates a real hiring risk, not a hypothetical one.
Beyond those two groups, anyone taking calls where trust and money intersect, journalists verifying sources, small business owners closing deals over video, or solo founders fundraising remotely, gets a reasonable case for it too. Someone who mostly uses video calls to catch up with family has a much weaker one.
What Halo cannot do yet
It is Windows only. macOS support is described as “in development” with no committed date. If you are on a Mac, your only option right now is emailing Scam.ai to register interest in the preview.
It only checks faces you can see. Halo detects synthetic video, not synthetic audio. A voice-cloning attack with no video component, like a phone call or an audio-only conference line, is outside what Halo currently covers. Scam.ai has a separate partnership with Modulate to add synthetic voice detection to its broader platform, but that integration currently sits in Scam.ai’s enterprise API, not inside the Halo consumer app.
It is one signal, not a verdict. Scam.ai is explicit in its own FAQ that Halo gathers evidence across multiple frames before changing a verdict, which is a sensible design choice against false positives, but it also means a very short appearance or a heavily compressed feed may not get flagged before the call ends. Treat an alert as a strong reason to stop and verify through a second channel, like calling the person back on a known number, not as the only check you will ever need.
The fraud numbers in the launch materials need a little unpacking. Scam.ai’s press release cites the FBI’s 2025 IC3 report and attributes $2.9 billion in losses to “business email compromise and deepfake-enabled wire fraud” combined. The actual 2025 IC3 report puts total business email compromise losses at $3.05 billion for the year, and separately breaks out AI-related crime as a new category for the first time: over 22,000 complaints and $893 million in confirmed AI-assisted losses across every crime type, not just BEC. Of that $893 million, the FBI attributes roughly $30 million specifically to AI-assisted business email compromise. The $2.9 billion figure in Scam.ai’s release rounds close to the full BEC number, not the much smaller slice the FBI actually confirms as AI-driven. The direction of the trend is real, and the FBI itself calls its AI figure a floor, not a ceiling, since most victims cannot tell whether AI was involved at all. The specific dollar figure just deserves more precision than the press release gives it.
How it compares to other deepfake detection tools
Halo occupies a specific spot in a market that has grown crowded fast. Most competitors are built for enterprise security teams, not for someone installing an app on their own laptop.
| Tool | What it is actually for | How it is deployed |
|---|---|---|
| Halo (Scam.ai) | Live video call protection for individuals and small teams | Desktop app, on-device, subscription |
| GetReal Security | Real-time identity and meeting protection for enterprises | Enterprise deployment, sales-led |
| Reality Defender | Multimodal detection (video, audio, image) at onboarding or upload | API, enterprise contract |
| Intel FakeCatcher | Physiological-signal video analysis for high-volume screening | Server-based, runs on Xeon hardware |
| Pindrop | Voice-clone detection for call centers | API, enterprise contract |
The practical difference is that almost everything else on that list is sold to a security team with a procurement budget. Halo is the rare option a single person can download, install, and start using in the same afternoon, which is exactly the gap Scam.ai is targeting.
What is coming next
Three things worth tracking. A macOS preview is in development with no announced date, so Mac users are on a waitlist for now. Voice detection through the Modulate partnership currently lives in Scam.ai’s enterprise API rather than inside Halo, but it is a reasonable bet it eventually reaches the consumer app given how much of this fraud combines a fake face with a cloned voice. And platform support is explicitly described as expanding beyond the current 10-plus apps, though Scam.ai has not said which ones come next.
The honest summary
Halo solves a real problem in a sensible way: check the call while it is happening, keep the video on the device, and get out of the way until something actually looks wrong. The pricing is reasonable for anyone who regularly approves payments or interviews candidates over video, and steep for anyone who does not. The Windows-only limitation will frustrate a chunk of its likely audience, and the launch materials oversell the fraud statistics a little more than the underlying FBI data supports.
None of that changes the core logic. A deepfake video call is convincing precisely because it exploits the same trust cues a real call would. Having something quietly watching for the signs you would miss is a reasonable trade for $15 a month, particularly if your job involves the words “confidential transaction.”
A finance employee at Arup lost $25.6 million because he trusted the faces on his screen. Fifteen dollars a month to make sure that is never you is one of the easier calls you will make this year.
Sources
- Scam.ai Launches Halo, Real-Time On-Device Deepfake Detection for Video Calls, Business Wire, August 5, 2026
- Scam.ai Announces Qualcomm Partnership, Launches Halo Deepfake Detection Model at Computex 2026, Business Wire, June 2026
- Halo product page, Scam.ai
- Scam.ai launches on-device deepfake detection for video calls, Biometric Update
- 2025 Internet Crime Report, FBI Internet Crime Complaint Center (IC3)
- ‘By 2028, one in four candidate profiles will be fake,’ Personnel Today
- Arup deepfake scam case background, CNN Business
Just Being Resourceful! 
Leave a comment