Google DeepMind Is Bringing AI Scam Detection to Your Phone

Morning, {{first name | folks}}! Google is bringing scam detection onto your phone, OpenAI is cutting malware analysis from weeks to hours, and Nvidia is putting $250 million behind AI safety research. Plus, Grok’s latest voice model is topping benchmarks while Anthropic expands its presence in Singapore. Let’s get into it.

Google DeepMind and Singapore’s GovTech have developed an AI model that spots scam attempts by analyzing conversations directly on a phone. It’s designed to catch impersonation and social engineering before people are tricked into handing over money or sensitive information.

The model runs on-device, so private conversations don’t need to be sent to the cloud for analysis. That could make scam detection more immediate and accessible, while keeping sensitive chats on the user’s phone.

OpenAI says its Daybreak system is helping SentinelOne analyzes complex malware in around eight hours, compared with the four to six weeks the work could take a senior researcher. That gives security teams a faster way to understand how malicious software works and investigate threats before they spread further.

The appeal isn’t just speed. AI can help more security analysts tackle work that once required highly specialized expertise, giving teams more capacity to investigate threats. Human researchers still need to verify findings and make the critical decisions, but they can spend less time on the initial analysis.

Artificial Analysis ranks SpaceXAI’s Grok Voice 3.0 first on its Speech-to-Speech Index, with a score of 86.9, ahead of the GPT and Gemini models tested. It also achieved a 94.8% task success rate and scored 98.5% on Big Bench Audio reasoning, edging out the evaluated Gemini model.

The results come with a trade-off: Grok Voice 3.0 ranks third in user preference, despite leading several performance tests. It’s available through SpaceXAI’s Realtime API for $0.08 per minute, with an average time to first audio of 0.81 seconds.

Anthropic has opened its fifth Asia-Pacific office in Singapore and signed an agreement with the government covering AI safety, model evaluation, and responsible adoption. Singapore ranks second globally for Claude usage per capita, at nearly six times what its population size would predict.

The expansion builds on existing adoption: more than 1,500 government engineers and technical staff use Claude Code, while Grab’s AI sales assistant saves around 40,000 hours every quarter. Anthropic will also work with Singapore on AI skills, security testing and helping businesses adopt AI responsibly.

Nvidia is committing $250 million in computing resources over five years to help UK researchers make AI agents safer and more secure. The work will involve universities including Oxford, UCL, and Imperial College London, with projects focused on testing AI systems, finding security weaknesses and improving safeguards.

As AI agents take on more tasks with less human input, testing how they behave outside controlled environments is becoming increasingly important. Nvidia’s investment gives researchers more computing power to study those risks before these systems are used more widely.