Scam.ai says Halo is now available for Windows as a local, real-time warning layer against deepfaked faces in Zoom, Microsoft Teams, and Google Meet calls. The practical appeal is clear: a system-tray application that watches a meeting window and alerts the user before a fraudulent “CFO” call turns into a wire transfer. But the company’s public product page shows Halo remains a beta access product, and the evidence published so far does not establish how accurately it detects current face-swap tools, which Windows PCs can run it, or how frequently it will mistakenly flag legitimate participants.
The August 5 announcement, issued through Business Wire, says Halo was developed with Qualcomm and runs its model on a Snapdragon neural processing unit rather than sending meeting content to a cloud service. Scam.ai’s own privacy policy supports the core privacy claim: it says Halo captures frames from a meeting window selected by the user, analyzes them locally, and keeps captured frames, participant scores, alerts, and saved evidence in the local Windows user profile.
That architecture is meaningful for IT teams handling sensitive calls. It avoids creating a new repository of board meetings, employee interviews, financial approval calls, or customer conversations on a third-party detection vendor’s servers. It also removes cloud round-trip latency from the critical moment: the user needs a warning before approving a payment, sharing credentials, or accepting a claimed identity.
What Halo does not yet provide is the independent evidence required to treat a detection alert as a security control rather than an additional signal.
Scam.ai previously announced Halo and its Qualcomm relationship on June 25 at Computex 2026, describing the product as available in June and optimized for Qualcomm-powered desktop computers. The August announcement recasts that launch around Windows availability and Snapdragon NPU inference.
However, Scam.ai’s current Halo page says the product is in beta mode and directs prospective users to apply for access. The site presents no public installer, no published price, and no enterprise licensing terms. Its German-language product page goes further, describing the offering as early access and a waitlist.
For administrators, that makes this an evaluation product rather than a broadly deployable endpoint control. A beta application that captures meeting windows needs review for software distribution, endpoint detection and response interactions, privacy notices, local-data retention, and supportability before it can be installed across finance, recruiting, executive-assistant, or legal teams.
Scam.ai also has not published a device compatibility list. The August release specifically emphasizes Snapdragon NPUs, while the privacy policy says Halo analyzes meeting frames using the computer’s “NPU/CPU.” That wording leaves unresolved whether Windows PCs based on Intel Core Ultra, AMD Ryzen AI, conventional x86 hardware without an NPU, or Windows on Arm devices other than a defined Snapdragon set are supported—and if so, at what performance level.
The company’s June announcement described Halo as optimized for Qualcomm-powered PCs, not exclusive to them. Until Scam.ai publishes supported processor families, minimum RAM and Windows-version requirements, and expected CPU utilization, IT departments should not assume a standard Windows fleet can run it without performance or battery tradeoffs.
That is a narrower task than authenticating a person, proving an executive approved a transfer, or detecting all AI-generated meeting content. The company’s materials do not say Halo verifies a speaker’s identity against an enrolled biometric profile. They do not claim it analyzes audio, detects voice cloning, validates meeting invitations, identifies compromised accounts, or checks the destination account of a requested payment.
Those distinctions are central to the type of fraud Scam.ai is using to sell the product. In the 2024 Arup case, confirmed by Arup and reported by The Guardian, a Hong Kong employee was deceived into making transfers totaling about $25 million after an AI-generated video call impersonated senior personnel. The visual deception mattered, but the incident also depended on a payment process that allowed 15 transfers to be authorized after the call.
A face-detection alert could give an employee reason to stop. It cannot replace a payment control that independently verifies new beneficiaries, requires out-of-band confirmation through trusted contact information, separates initiation from approval, and holds unusual transfers for a second reviewer. A deepfake detection system belongs in the challenge-and-verify stage of a transaction process, not at the point where a company decides a video call is sufficient authorization.
The company’s framing is also more specific than its press release suggests. Halo may observe a meeting application without a Teams, Zoom, or Meet plugin, which reduces integration friction, but it cannot see what is outside the selected window. That includes a phone call running alongside the meeting, a second monitor, a chat-based payment request, or voice-only impersonation.
The August announcement says Halo processes “each video frame” as a call happens. The company’s own product page instead describes a sampling rate of about four scans per second. Those claims may be compatible if the software samples frames at intervals, but Scam.ai has not documented the pipeline well enough to establish exactly what it means by real time.
More consequentially, there is no published accuracy figure, false-positive rate, false-negative rate, test corpus, model card, benchmark methodology, or breakdown by conferencing codec, lighting, camera quality, bandwidth reduction, makeup, virtual backgrounds, screen-shared video, or accessible-video accommodations. No independent lab, university, enterprise customer, or security research outlet appears to have published validation of Halo’s detection performance as of August 5.
That omission is the major operational limitation. Deepfake detectors face a difficult asymmetric problem: a false negative can let an impersonation proceed, while a false positive can derail a legitimate executive or candidate interview. A warning tool can be valuable even with imperfect accuracy, but security leaders need measured error rates to decide whether it is appropriate for high-stakes workflows and what escalation process should follow an alert.
Scam.ai’s public materials also do not state how alerts are presented to meeting participants, whether the application can distinguish a real camera feed from replayed video, whether it detects partial face swaps, or what happens when the target’s face is briefly off camera. These are not edge cases in video conferencing; they are normal conditions.
That is not a contradiction of the “no meeting media upload” statement. It does mean “fully on-device” should be read as a description of media analysis, not as a claim that Halo is an entirely offline executable with no service dependencies.
The same policy says captured frames, detection results, participant scores, and any saved evidence remain in the local user profile under
For regulated teams, the local audit trail is both an advantage and a new records-management question. It can support an investigation after a suspected impersonation attempt, but it may also contain sensitive indications about who appeared on a meeting and when an alert was generated.
The 2025 IC3 report lists Business Email Compromise losses at approximately $3.05 billion, based on 24,768 complaints. It separately records “AI Related” as a descriptor attached to 22,364 complaints across crime categories. The report does not provide a total specifically for deepfake-enabled wire fraud, nor does it combine that subset with BEC into a $2.9 billion figure.
The correction does not diminish the threat. If anything, the official figure shows that BEC remains a multibillion-dollar fraud category, while the Bureau’s AI-related descriptor indicates that AI has become a reportable dimension across multiple scam types. But calling the number a deepfake-wire-fraud total gives the statistic a precision the FBI has not reported.
Halo addresses a real gap between identity controls at login and the human trust decision made during a call. The product’s local processing model is well suited to privacy-sensitive Windows environments, particularly if Scam.ai can demonstrate reliable detection across ordinary business hardware. For now, its beta status, undocumented hardware support, absent accuracy data, and lack of independent validation mean it should be tested as a supplementary fraud-warning tool—not approved as proof that the face on a Teams or Zoom call is genuine.
That architecture is meaningful for IT teams handling sensitive calls. It avoids creating a new repository of board meetings, employee interviews, financial approval calls, or customer conversations on a third-party detection vendor’s servers. It also removes cloud round-trip latency from the critical moment: the user needs a warning before approving a payment, sharing credentials, or accepting a claimed identity.
What Halo does not yet provide is the independent evidence required to treat a detection alert as a security control rather than an additional signal.
Halo’s public availability is narrower than “available now” suggests
Scam.ai previously announced Halo and its Qualcomm relationship on June 25 at Computex 2026, describing the product as available in June and optimized for Qualcomm-powered desktop computers. The August announcement recasts that launch around Windows availability and Snapdragon NPU inference.However, Scam.ai’s current Halo page says the product is in beta mode and directs prospective users to apply for access. The site presents no public installer, no published price, and no enterprise licensing terms. Its German-language product page goes further, describing the offering as early access and a waitlist.
For administrators, that makes this an evaluation product rather than a broadly deployable endpoint control. A beta application that captures meeting windows needs review for software distribution, endpoint detection and response interactions, privacy notices, local-data retention, and supportability before it can be installed across finance, recruiting, executive-assistant, or legal teams.
Scam.ai also has not published a device compatibility list. The August release specifically emphasizes Snapdragon NPUs, while the privacy policy says Halo analyzes meeting frames using the computer’s “NPU/CPU.” That wording leaves unresolved whether Windows PCs based on Intel Core Ultra, AMD Ryzen AI, conventional x86 hardware without an NPU, or Windows on Arm devices other than a defined Snapdragon set are supported—and if so, at what performance level.
The company’s June announcement described Halo as optimized for Qualcomm-powered PCs, not exclusive to them. Until Scam.ai publishes supported processor families, minimum RAM and Windows-version requirements, and expected CPU utilization, IT departments should not assume a standard Windows fleet can run it without performance or battery tradeoffs.
The product detects faces, not the whole impersonation attack
Halo’s own product page describes the tool as flagging “synthetic faces and faceswaps” during live calls. Its workflow is window capture: it scans faces visible in the chosen meeting application, then produces an alert if a face appears synthetic.That is a narrower task than authenticating a person, proving an executive approved a transfer, or detecting all AI-generated meeting content. The company’s materials do not say Halo verifies a speaker’s identity against an enrolled biometric profile. They do not claim it analyzes audio, detects voice cloning, validates meeting invitations, identifies compromised accounts, or checks the destination account of a requested payment.
Those distinctions are central to the type of fraud Scam.ai is using to sell the product. In the 2024 Arup case, confirmed by Arup and reported by The Guardian, a Hong Kong employee was deceived into making transfers totaling about $25 million after an AI-generated video call impersonated senior personnel. The visual deception mattered, but the incident also depended on a payment process that allowed 15 transfers to be authorized after the call.
A face-detection alert could give an employee reason to stop. It cannot replace a payment control that independently verifies new beneficiaries, requires out-of-band confirmation through trusted contact information, separates initiation from approval, and holds unusual transfers for a second reviewer. A deepfake detection system belongs in the challenge-and-verify stage of a transaction process, not at the point where a company decides a video call is sufficient authorization.
The company’s framing is also more specific than its press release suggests. Halo may observe a meeting application without a Teams, Zoom, or Meet plugin, which reduces integration friction, but it cannot see what is outside the selected window. That includes a phone call running alongside the meeting, a second monitor, a chat-based payment request, or voice-only impersonation.
The “real-time” claim needs performance data, not just local inference
Scam.ai says Halo analyzes video locally and surfaces a warning while the call is live. Its product page says it scans roughly four times per second. That is enough to produce a near-real-time indication for an on-screen face, but it is not the same as analyzing every video frame at typical conferencing frame rates.The August announcement says Halo processes “each video frame” as a call happens. The company’s own product page instead describes a sampling rate of about four scans per second. Those claims may be compatible if the software samples frames at intervals, but Scam.ai has not documented the pipeline well enough to establish exactly what it means by real time.
More consequentially, there is no published accuracy figure, false-positive rate, false-negative rate, test corpus, model card, benchmark methodology, or breakdown by conferencing codec, lighting, camera quality, bandwidth reduction, makeup, virtual backgrounds, screen-shared video, or accessible-video accommodations. No independent lab, university, enterprise customer, or security research outlet appears to have published validation of Halo’s detection performance as of August 5.
That omission is the major operational limitation. Deepfake detectors face a difficult asymmetric problem: a false negative can let an impersonation proceed, while a false positive can derail a legitimate executive or candidate interview. A warning tool can be valuable even with imperfect accuracy, but security leaders need measured error rates to decide whether it is appropriate for high-stakes workflows and what escalation process should follow an alert.
Scam.ai’s public materials also do not state how alerts are presented to meeting participants, whether the application can distinguish a real camera feed from replayed video, whether it detects partial face swaps, or what happens when the target’s face is briefly off camera. These are not edge cases in video conferencing; they are normal conditions.
Local media processing does not mean zero network activity
Scam.ai’s claim that meeting video and audio are not sent to the cloud is broadly supported by its privacy policy, with an important qualification. The policy says Halo requires sign-in, using an identity service to verify the user and issue a session token that unlocks on-device detection. It also offers optional anonymous crash and performance diagnostics through Sentry, disabled by default.That is not a contradiction of the “no meeting media upload” statement. It does mean “fully on-device” should be read as a description of media analysis, not as a claim that Halo is an entirely offline executable with no service dependencies.
The same policy says captured frames, detection results, participant scores, and any saved evidence remain in the local user profile under
%LOCALAPPDATA%. It says uninstalling Halo may leave that local folder behind depending on how the user removes the application. Organizations evaluating the tool should therefore determine who can access those local records, whether they are encrypted at rest, how long evidence persists, whether endpoint backups collect it, and whether a removal script needs to delete residual data.For regulated teams, the local audit trail is both an advantage and a new records-management question. It can support an investigation after a suspected impersonation attempt, but it may also contain sensitive indications about who appeared on a meeting and when an alert was generated.
The FBI statistic in the launch announcement does not match the FBI report
Scam.ai’s announcement says the FBI’s 2025 IC3 report attributed an estimated $2.9 billion in losses to “business email compromise and deepfake-enabled wire fraud.” The FBI report does not publish that combined figure.The 2025 IC3 report lists Business Email Compromise losses at approximately $3.05 billion, based on 24,768 complaints. It separately records “AI Related” as a descriptor attached to 22,364 complaints across crime categories. The report does not provide a total specifically for deepfake-enabled wire fraud, nor does it combine that subset with BEC into a $2.9 billion figure.
The correction does not diminish the threat. If anything, the official figure shows that BEC remains a multibillion-dollar fraud category, while the Bureau’s AI-related descriptor indicates that AI has become a reportable dimension across multiple scam types. But calling the number a deepfake-wire-fraud total gives the statistic a precision the FBI has not reported.
Halo addresses a real gap between identity controls at login and the human trust decision made during a call. The product’s local processing model is well suited to privacy-sensitive Windows environments, particularly if Scam.ai can demonstrate reliable detection across ordinary business hardware. For now, its beta status, undocumented hardware support, absent accuracy data, and lack of independent validation mean it should be tested as a supplementary fraud-warning tool—not approved as proof that the face on a Teams or Zoom call is genuine.
References
- Primary source: 01net
Published: 2026-08-05T23:09:00+00:00
Loading…
www.01net.it - Related coverage: arstechnica.com
Loading…
arstechnica.com - Related coverage: markets.financialcontent.com
Loading…
markets.financialcontent.com - Related coverage: theguardian.com
Loading…
www.theguardian.com - Related coverage: cyberhelmets.com
Loading…
cyberhelmets.com - Related coverage: missionmedia.asia
Loading…
missionmedia.asia - Related coverage: ironhackers.es
Loading…
ironhackers.es - Related coverage: valtikstudios.com
Loading…
www.valtikstudios.com - Related coverage: timewell.jp
Loading…
timewell.jp - Related coverage: techradar.com
Loading…
www.techradar.com - Related coverage: scam.ai
Loading…
scam.ai - Related coverage: qualcomm.com
Loading…
www.qualcomm.com - Related coverage: scam.ai
Loading…
www.scam.ai - Related coverage: finance.yahoo.com
Loading…
finance.yahoo.com - Related coverage: natlawreview.com
Loading…
natlawreview.com - Related coverage: streetinsider.com
Loading…
www.streetinsider.com - Related coverage: ca.marketscreener.com
Loading…
ca.marketscreener.com - Related coverage: markets.chroniclejournal.com
Loading…
markets.chroniclejournal.com - Related coverage: winzheng.com
Loading…
www.winzheng.com