Scammers are getting an AI-shaped headache. Google DeepMind and Singapore’s Government Technology Agency, or GovTech, announced an on-device scam detection model on October 14, 2026, designed to warn smartphone users about manipulative conversations without sending private messages to the cloud.
The model processes conversation data on the device itself, a design that keeps message content away from cloud systems during detection. That matters in Singapore, where scams make up over half of reported crimes and have caused more than S$4 billion in losses since 2019.
Up to 80% of all scam losses come from social engineering tactics, which manipulate people into sending money or revealing information. The model targets that human weakness by examining conversations for signs of manipulation rather than waiting for a transaction to expose the fraud.
A two-stage defense for private conversations
Google DeepMind fine-tuned its lightweight Gemma model on local scam conversations in Singapore. The system uses a two-stage pipeline: a lightweight classifier handles the first pass, then the Gemma model performs a contextual evaluation.
This split gives the system a faster first filter before asking the larger model to examine the conversation’s meaning. The approach also keeps the processing on the smartphone, avoiding the familiar trade-off in which privacy becomes the bill for automated protection.
Google DeepMind and GovTech presented the collaboration at Singapore International Cyber Week 2026. The event’s opening address was delivered on October 13, 2026, by K Shanmugam, Singapore’s Senior Minister and Minister for Home Affairs.
From detecting scams to wasting scammers’ time
The Singapore announcement arrived alongside a wider set of law enforcement and technology efforts. On October 13, 2026, the Justice Department and FBI announced an operation against West African criminal scam networks, supported by law enforcement agencies from Australia, Canada, India, New Zealand, and Interpol.
The operation led to dozens of arrests and rescued 117 trafficking victims from scam compounds in Ghana. Authorities in Ghana seized hundreds of electronic devices and identified at least $11 million in potentially freezable funds.
Meta removed more than 3.7 million accounts, pages, or groups accused of targeting users in the United States and United Kingdom, with most profiles impersonating American and British users. Microsoft suspended accounts suspected of being linked to cybercrime, while Coinbase helped identify more than 65 victims and stopped more than $415,000 in potential scams.
Meta said, “Victims had been forced to work for scam operations by criminal syndicates and were rescued from imprisonment in a residential compound.” The victims were malnourished and subjected to physical and psychological abuses — a useful reminder that “online fraud” can describe an offline prison.
Apate, an Australian company, is taking the opposite approach from detection: its system keeps scammers talking. The company has spent two years building a platform that diverts phone scammers onto calls with AI bots, and its system has around 350,000 bots that answer calls, infiltrate scam chat groups online, and respond to text messages.
Apate’s bots use different personalities, language skills, and profiles to give scammers enough hope to stay on the phone. The platform, used by banks and supported by telecom companies, has collected more than 250,000 pieces of information about fraudsters in real time.
Kernel Panic tested Apate’s tool with a demo version and found the system engaged and responsive during testing. The AI did not persuade users to invest in a cryptocurrency opportunity after six minutes of effort, but that was not the point: its purpose is to keep scammers occupied while gathering intelligence about their operations.
Cybercriminals initiate billions of messages and calls each year, and law enforcement and research efforts have not stopped the expansion of digital scamming overall. AI platforms are now being used to scam the scammers by tricking them into talking to lifelike bots.
Open source honeypot providers have also been incorporating large language models into their systems. Mark Vero, a doctoral researcher at ETH Zurich, is associated with the finding that an LLM-powered honeypot keeps AI agents attacking for longer than a predictable honeypot.
“Fraud is a global problem and it takes partnership working across international boundaries to tackle it,” said Nick Sharp, NCA deputy director of fraud. The emerging strategy is clear: detect manipulation on the device, disrupt criminal networks through coordinated operations, and make automated attackers waste their time on machines that never send the money.
Based on
