Android 17 launches with new multitasking tools as Google expands Gemini features

Title: AI Mainstream Push Meets Safety as Models Proliferate

01Android 17 Brings Gemini Deeper Into the Smartphone

Google has released Android 17 alongside Wear OS 7, embedding its Gemini AI features deeper into the mobile experience. The update introduces new multitasking tools, expanded parental controls, and enhanced security features designed to bring AI capabilities to billions of users. A simultaneous Pixel Drop delivers Google's latest AI models directly to Pixel devices, signaling a coordinated push to make artificial intelligence a core component of the smartphone experience rather than a separate application.

The launch represents Google's most aggressive integration of AI into its mobile operating system to date. By embedding Gemini across system functions—from calendar management to messaging—Google is betting that users will expect AI assistance as fundamental as cellular connectivity. Industry analysts note the timing coincides with intensifying competition from Samsung and Apple, both of which have expanded their own AI offerings in recent months.

02'Dangerous' AI Models Inevitable Despite Crackdowns

A Wired investigation reveals the US government crackdown on Anthropic's Claude Fable 5 and Mythos 5 models exposes a fundamental reality: AI systems with concerning capabilities will proliferate regardless of policy interventions. The analysis argues that advanced hacking capabilities in AI models will soon become the norm, rendering current regulatory approaches insufficient. The piece suggests that attempts to restrict frontier models only delay their inevitable emergence through alternative channels.

The investigation highlights a growing tension between safety-focused AI developers and government agencies seeking to limit model capabilities. Anthropic has publicly defended its work, arguing that responsible deployment with safety measures is preferable to driving capability development underground. The Wired analysis concludes that policymakers face a choice between engaging constructively with leading AI labs or risk creating a fragmented landscape where safety considerations are abandoned entirely.

03OpenAI Tests Model Safety With Deployment Simulation

OpenAI has introduced Deployment Simulation, a method to predict AI model behavior before public release using real conversation data to improve safety evaluation accuracy. The technique simulates how models will perform in actual deployment conditions, allowing researchers to identify potential failure modes before they affect users at scale. The approach represents a significant shift from traditional benchmark-based evaluation toward continuous, real-world testing frameworks.

The company says Deployment Simulation addresses a critical gap in current safety research: the difficulty of anticipating how models will behave when exposed to diverse user interactions. By modeling deployment conditions, OpenAI aims to catch safety issues earlier in the development pipeline. The method has already been applied to recent model releases, though the company has not disclosed specific findings or how the simulations have influenced development decisions.

Also today

04SpaceX to acquire Cursor for $60B in stock, days after blockbuster IPO — SpaceX's acquisition of AI coding tool Cursor targets a claimed $26T addressable market, just days after the SpaceX IPO.
05Anthropic's feud with Trump admin may actually help it, sales data suggests — Ramp data indicates Anthropic's business growth is so strong its government conflict may boost rather than harm sales.
06ChatGPT's market share slips below 50% for first time — OpenAI's ChatGPT fell below 50% market share while retaining the top spot with 1.1B monthly users, as Gemini (662M) and Claude (245M) gain ground.
07Critical Copilot vulnerability allowed hackers to steal 2FA codes — Microsoft's Copilot suffered a SearchLeak exploit enabling 2FA code theft, illustrating persistent failures in LLM security.
08Probably raises $9M to build a more reliable kind of AI — AI startup Probably secured seed funding to develop models that prevent hallucinations and achieve accuracy comparable to deterministic systems.
09Georgi Gerganov vouches for Qwen3.6-27B as capable local coding model — Developer Georgi Gerganov confirms Qwen3.6-27B runs well on M2 Ultra and RTX 5090 hardware for local coding tasks.
10Running local models has reached a practical threshold — A developer reports local AI models have reached a practical threshold for real work, gaining significant traction with 926 Hacker News points.