Halo by Scam AI is an on-device deepfake detector that Scam.ai built with Qualcomm to watch live video calls and flag synthetic faces while the call is still happening. It shipped in June 2026, runs locally on Qualcomm-powered desktops, and is aimed at HR teams, recruiters, and executives who take high-stakes calls from strangers.
Short answer: Halo by Scam AI is an on-device model that scans live video calls — Zoom, Teams, Meet — for synthetic or AI-generated faces in real time, without sending any video to the cloud. It launched in June 2026 through a Qualcomm partnership, runs only on Qualcomm-powered desktops today, and Scam.ai hasn't published a standalone price for it; access goes through an enterprise conversation.
In my testing, I couldn't run Halo itself — it requires a Qualcomm-powered desktop and, as of this writing, ships through enterprise deals rather than a public download — so I worked from Scam.ai's own site, its pricing page, and the Qualcomm/Computex 2026 press materials, checking every figure below on July 31, 2026. What follows is what Halo actually does, how the underlying Scam.ai detection platform prices out if you want to test the technology today, and where a background deepfake watcher is worth deploying versus overkill.
What you'll need
Halo only runs on desktop computers built around Qualcomm processors — there's no version for Intel or AMD machines yet, and no phone or laptop-webcam-only mode. You'll also need to be the kind of organization Scam.ai is actually selling to: it's pitched at HR and recruiting teams doing remote interviews, plus executives like CFOs and CEOs who take frequent high-stakes calls from people they've never met in person. There's no self-serve signup for Halo itself. If you just want to test Scam.ai's underlying detection technology on images or video clips rather than live calls, that part does have a free tier — 200 images a month on the Eva-v1-Fast model, no card required.
Step-by-step: getting access to Halo by Scam AI
1. Confirm your hardware
Check that the machines you'd deploy this on run Qualcomm silicon. Scam.ai and Qualcomm built Halo specifically for that hardware, and as of the June 2026 launch it isn't available on other chipsets.
2. Try the underlying detection engine first
Before committing to an enterprise conversation about Halo, sign up for Scam.ai's Self-Serve plan and run a few real images or video clips through the Eva-v1-Fast model. It's free for the first 200 images a month, then $0.05 per image, and it's the same core detection technology Halo runs on-device.
3. Bring a real use case to the conversation
Halo doesn't have a published price or a self-serve rollout, per Scam.ai’s pricing page. Come with a specific problem — screening video interviews, protecting exec video calls — rather than general interest in deepfake detection.
4. Pilot on the interviewers or executives most exposed
Deploy Halo first on the desktops of the people fielding the highest volume of first-time video calls with strangers: recruiters running first-round interviews and finance leaders who approve wire transfers over video.
5. Treat a flag as a prompt to verify, not a verdict
When Halo flags a call, the right next step is to verify identity through a second channel — a callback to a known number, a badge check — not to assume fraud outright. On-device, real-time detection is a tripwire, not a courtroom.
6. Reassess after your renewal fraud numbers come in
Because Halo is new and enterprise-priced, judge it against a real metric: fraudulent-interview attempts caught, or wire-transfer requests halted, over a full quarter — not against the pitch deck.
Questions worth copying into your vendor call
Since Halo doesn't take prompts the way a chat tool does, the copy-paste value here is what to actually ask on the sales call, based on what Scam.ai has and hasn't published:
- "Which of our current video platforms — Zoom, Teams, Meet — has Halo actually been tested against in production, not just internally?"
- "What happens on non-Qualcomm hardware — is there a roadmap, or is this a hard requirement indefinitely?"
- "Walk us through a false positive: what does a flagged-but-real caller see, and how fast can they clear it?"
- "Is any video or metadata retained anywhere, or does 'runs on-device' mean nothing leaves the machine at all?"
- "How does pricing compare to just running our recorded interviews through the Eva-v1-Pro forensic API after the fact?"
Common mistakes to avoid
The biggest one is assuming Halo replaces identity verification rather than adding a layer to it — a flagged call still needs a human to confirm who's actually on the line. Second, buying Qualcomm hardware specifically for Halo before confirming your existing desktops already qualify; check first. Third, rolling this out company-wide on day one instead of piloting it with the recruiters and executives most exposed to deepfake calls, since it's an enterprise deal with real cost attached. Fourth, skipping the free Self-Serve tier on Scam.ai's core API — testing detection accuracy on your own sample images before an enterprise conversation gives you a real basis for judging the pitch. Fifth, treating a real-time flag as proof of fraud instead of a reason to verify through a second channel; on-device detection is fast but not infallible.
Tools that make this easier
Halo is one narrow piece of a bigger "AI trust and safety" category, and it's worth seeing where it sits next to tools you may already use. Lunen.ai tackles a related but broader problem — governing what AI agents are allowed to touch inside a company — if identity fraud on calls is one symptom of a wider AI-risk gap you're managing. If Halo's target use case is screening job candidates, best AI tool for job searching is worth reading from the other side of that interview. For a look at how convincing synthetic voice has gotten, which is half of what makes deepfake video calls dangerous, see how to use ElevenLabs. If you're evaluating whether a given AI system's underlying capability is actually a risk to your organization, my writeup on the UK AISI and CAISI assessment of Kimi K3’s cyber capabilities shows how independent testers score that kind of claim. Skippr AI is a useful contrast too — it embeds an agent inside your own product rather than watching for fraud from outside it. And I test every tool on this site the same way, hands-on against the vendor's own site and pricing page — how we test AI tools explains that process in full.
Halo versus the rest of Scam.ai's detection stack
| Option | Price | When it runs | Best for |
|---|---|---|---|
| Halo (on-device, live) | Not published — enterprise only | During the call, in real time | HR interviews and exec calls on Qualcomm desktops |
| Eva-v1-Fast (Self-Serve API) | Free for 200 images/mo, then $0.05/image | On-demand, per image or clip | Testing detection accuracy before an enterprise deal |
| Eva-v1-Pro (Enterprise API) | Custom, sales conversation | On-demand, forensic-grade in under 4 seconds | Legal-grade reports after an incident |
| No detection tool | $0 | Never | Organizations not yet worried about deepfake calls |
The Self-Serve and Enterprise API rows cover images and recorded clips; Halo is the only one built for a call that's still in progress.
My take
Deloitte's Center for Financial Services projects that generative-AI-enabled fraud could reach $40 billion in U.S. losses by 2027, up from $12.3 billion in 2023 (Deloitte, 2026), and Scam.ai cites its own figure that deepfake fraud attempts are up more than 2,000% over three years. Against that backdrop, a passive, on-device watcher for the video calls where fraud actually starts is a sane idea, not hype. What I'd flag before anyone signs: Halo is Qualcomm-hardware-only today, there's no public price, and Scam.ai itself says only 31% of HR leaders feel equipped to catch identity fraud in interviews — meaning the tool is solving a real, admitted gap rather than a manufactured one. If your team already runs Qualcomm desktops and takes enough stranger-to-stranger video calls that a single successful deepfake would hurt, it's worth the enterprise call. If you're a smaller team on standard laptops, start with the free Eva-v1-Fast tier on suspicious clips instead of chasing hardware you don't have.
Frequently Asked Questions
Is Halo by Scam AI free?
No. Halo itself has no published price and is sold through an enterprise conversation. Scam.ai's underlying detection API does have a free tier — 200 images a month on Eva-v1-Fast, then $0.05 per image.
How long does it take to get Halo running?
There's no instant signup. You need Qualcomm-powered desktop hardware and a sales conversation with Scam.ai; the company hasn't published a typical rollout timeline since Halo only launched in June 2026.
What's the easiest way to evaluate Halo before committing?
Run real sample images or clips through Scam.ai's free Self-Serve tier first. It's the same Eva-v1 detection model Halo runs on-device, so the accuracy you see there is a reasonable preview.
Does Halo record or store the video it analyzes?
According to Scam.ai, Halo runs entirely on the device with nothing recorded and nothing sent off the machine — detection happens locally during the call.
Does Halo work on any computer?
No. As of its June 2026 launch, Halo is built specifically for Qualcomm-powered desktop computers; there's no Intel, AMD, or mobile version yet.