The honest view from inside technical leadership
AI News

AI News

Daily AI news with an engineering leadership take, each story credited to its original AI News report.

Security · AI News

Singapore’s On-Device Scam Detector Keeps Messages on the Phone, and Its First App Can Only Warn

Google DeepMind detailed the model in October, and GovTech has since shipped the first app built on the collaboration. Its privacy promise rests on where the model runs, which is a design choice engineering teams can copy.

Line-art illustration of a smartphone with a chat bubble and a warning triangle beside a small chip, representing an on-device scam detector

GovTech Singapore and Google DeepMind have built an on-device scam detector: a fine-tuned version of Google’s Gemma model that analyses conversations on the phone, so message content does not have to leave the device. Google DeepMind described the model on October 14, 2026, and GovTech launched AntiScam Pango, a beta Android app whose AI is supported by the same collaboration, on November 4. The app warns users about suspicious conversations and cannot block or delete anything.

Social Engineering Now Drives Up to 80% of Singapore’s Scam Losses

According to Google DeepMind, scams make up more than half of Singapore’s reported crimes and have caused more than S$4 billion in losses since 2019. The tactics have shifted from simple phishing links to long manipulative conversations that end in a request for money, and DeepMind says those social engineering scams now account for up to 80% of losses. Generic models struggle with local deception, so the teams fine-tuned Gemma on local scam conversations, including impersonation of public institutions, banks and local figures. The work sits under the National AI Partnership between Google DeepMind and the Singapore government, and it was presented at Singapore International Cyber Week in October.

The On-Device Scam Detector Runs a Cheap Classifier Before It Calls Gemma

Battery life is the hard constraint on a phone, so the design uses two stages. A lightweight classifier screens every inbound message at low power. Only when it flags a possible threat does the fine-tuned Gemma model look at a larger snapshot of the conversation to confirm the scam or discard the alarm. DeepMind says fine-tuning took place in a GovTech-managed Google Cloud environment it calls a Sovereign Sandbox, and that training excluded private and personally identifiable data in favour of anonymised datasets drawn from scam reports and intelligence feeds. The model reads conversations only with user consent, and because it runs entirely on the device, DeepMind says neither Google nor the Singapore government can access those conversations through it.

AntiScam Pango Ships the Model’s Approach as a Warning-Only Beta

Jasmin Lau, Senior Minister of State for Digital Development and Information, announced AntiScam Pango at STACK Conference 2026. Once a user grants permissions, the app runs in the background on Android and reads incoming message notifications from WhatsApp, Telegram and SMS. It chains messages from the same sender into one conversation, so it can catch scams that build up over several exchanges, and it explains why a flagged conversation looks risky. GovTech says all analysis happens on the device and message content is sent to external servers only when a user chooses to report it. The release credits the app’s AI to the DeepMind collaboration without naming the model, and neither the release nor DeepMind’s post publishes accuracy or false-positive figures. GovTech plans to keep improving the app during the beta and will explore integrating it with ScamShield, its existing app for blocking scam calls and filtering suspicious SMS.

Why This Matters for Engineering Leadership

The Singapore project puts its privacy promise into the architecture. A policy promising that nobody reads your messages has to be taken on trust. A model that runs on the handset, with a network path that carries message content only when the user presses report, can be checked by anyone who inspects the traffic. That is the pattern to take from this story for any AI feature that touches sensitive content, whether it scans support chats, emails or code. First, ask whether the inference can run where the data already lives, on the endpoint or inside the customer’s tenancy, before designing a pipeline that copies the data to a central model. Second, use the two-stage shape to make that affordable: a cheap filter on everything, and the expensive model only on what the filter escalates. Third, scope the first release the way GovTech did, with warnings and no automatic action, until you have measured false positives on real traffic. Publish those numbers before you let the system block, delete or hold anything, because a confident warning is cheap to ignore and a wrong block is expensive to undo.

FAQ

What is Singapore’s on-device scam detector?

It is a scam detection model built by GovTech Singapore and Google DeepMind by fine-tuning Google’s Gemma on local scam conversations. It analyses messages on the phone instead of on a server.

How does the scam detection pipeline work?

A lightweight classifier screens incoming messages first. When it flags a possible threat, the fine-tuned Gemma model reviews a larger part of the conversation to confirm the scam or dismiss the alert.

What is AntiScam Pango?

AntiScam Pango is a beta Android app GovTech launched on November 4, 2026. It reads incoming notifications from WhatsApp, Telegram and SMS, with permission, and alerts users to conversations that look like scams.

Can AntiScam Pango block scam messages?

No. GovTech says the app cannot block or delete messages. It explains why a conversation looks risky and leaves the response to the user.

Does Google or the Singapore government see users’ messages?

Google DeepMind says the model runs entirely on the device, so neither the government nor Google can access conversations through it. GovTech says message content leaves the phone only if the user reports it.

Originally reported by Ryan Daws for AI News, October 15, 2026: GovTech Singapore and Google DeepMind detail on-device Gemma scam detector. Primary sources: Google DeepMind, “Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection” (October 14, 2026); GovTech Singapore, “GovTech Launches AntiScam Pango Beta to Help Users Spot Suspicious Scam Messages” (November 4, 2026).

Read more AI News