Reports

Mon, May 12 - Google's On-Device Gemini AI Cuts Scams by 80%

Hak

12 May 2025 — 3 min read

It's Monday, May 12, 2025, and you're reading the Agentive Daily Report.

Busy People's Section

Google debuts on-device Gemini AI in Chrome, blocking scams in real time and claiming an 80% reduction in scammy results.

OpenAI expands Stargate with global supercomputing hubs, launches custom enterprise fine-tuning, and offers stronger regional data controls.

Figma’s new AI tools let users build, code, and market websites within a single platform, mounting a credible challenge to Adobe and Canva.

Meta unveils open-source foundation models for perception and reasoning; its new smart glasses experiment with always-on 'super-sensing.'

Mistral releases Medium 3 model optimized for enterprise at lower costs, while Anthropic launches near real-time Web Search API for Claude.

Microsoft and others warn US Senate that fragmented regulation may cede AI dominance to China; OpenAI, CoreWeave, and AMD push for streamlined permitting.

Today's Top Stories

Chrome’s On-Device AI Raises the Bar for Consumer Safety

Google has fully embedded Gemini Nano into Chrome’s Enhanced Protection, bringing real-time local scam detection to both desktop and Android. By running the model on-device, Google avoids sending sensitive data to the cloud, which is a significant privacy win for users who are weary of remote monitoring. The early metrics—an 80% drop in scam search results and hundreds of millions of blocked attempts daily—show what practical, responsible LLM deployment can look like at an industrial scale. What’s notable here is the direct impact: fewer scams, less fraud, and more trust for billions of Chrome users worldwide. With on-device LLMs now feasible, expect this playbook to spread to everything from email clients to IoT security.

OpenAI’s Stargate: A Global AI Power Grab, Now with Local Roots

OpenAI’s Stargate strategy is doubling down on two things: regional sovereignty and customizability. By establishing AI supercomputing centres globally and rolling out data-residency options across Asia, OpenAI is positioning itself as a partner (not a threat) to national governments and regulated industries. Meanwhile, the roll-out of reinforcement fine-tuning for its enterprise models gives organisations the tools to create truly domain-specific AI, with early evidence of real performance gains in taxes, healthcare coding, and legal analysis. On Capitol Hill, OpenAI and its partners argue that nimble regulation and massive infrastructure are necessary to keep pace with China—a narrative likely to shape global AI policy debates for months.

Figma’s AI Suite Aims to Disrupt Creative Workflows

Figma isn’t just releasing shiny features—it’s going after the all-in-one creative stack with Sites, Make (AI-powered coding), Buzz, and Draw. By combining rapid site building, AI code generation via Claude 3.7, and content creation under one roof, Figma gives individuals and teams, especially those outside the enterprise core, a low-barrier alternative to the Adobe-WordPress-Canva axis. This is a real shift for makers: less context-switching and more rapid prototyping, at a subscription price designed to undercut legacy players. Watch closely—AI-driven consolidation of creative toolchains is just beginning, but Figma’s blend of vertical integration and AI accessibility is a blueprint others will chase.

Fast Forward

Meta’s Super-Sensing Glasses: Meta’s smart glasses prototype is pushing boundaries in privacy and utility, piloting multi-hour always-on “super-sensing” and facial recognition. With open-source perception models now in the wild, the next frontier is balancing innovation with safeguards.
Multimodal Model Security Gaps: Research from Enkrypt AI finds that Mistral’s multimodal models are vastly more likely to generate unsafe content when manipulated, highlighting a growing risk as image+text AIs proliferate—the industry must address prompt injection and media attacks head-on.
Mistral, Anthropic, and the AI Search Wars: Mistral’s Medium 3 model aims for cost-efficient, flexible enterprise deployments, while Anthropic’s new API lets Claude ingest live web data, further eroding the “closed” model advantage and ramping up search competition with Google.
Microsoft and OpenAI on Regulatory Brinkmanship: At the US Senate, Microsoft, OpenAI, AMD, and CoreWeave called for unified regulatory frameworks and faster permitting of critical AI infrastructure, warning that balkanized rules or unchecked export controls risk losing the global AI race.
AI as a Reputation Risk: Studies show employees who use AI are often perceived as less competent or lazier, regardless of impact, reminding business leaders to manage bias and communicate how AI adds, not subtracts, value to teams.

New Tools Discovered

remio: Local-first AI note taker & personal knowledge hub, focused on privacy and ease of personal data management.
Nelly: Complete no-code platform for building, managing, and sharing teams of custom AI agents.
Xyla AI: Smart assistant that synthesises and organises information across sources for research-heavy professionals.
Find My Papers: Simple home for finding and managing academic papers with integrated AI search.
Factory: Streamlines large-scale AI operations, emphasising automation and scalability for teams deploying multiple agents.

Discover more tools at Agentive.Directory

That's a wrap for today! Thank you for reading this report.

Have thoughts on today's edition? Hit reply and let us know what you're thinking. Or if you've discovered a cool AI tool we should feature, drop us a line.

Until tomorrow,
Hak from Agentive.Studio