Stay ahead on AI.
All in ONE place.

Anthropic is making auto mode the default in Claude Code for Pro, Max, and Team plans starting August 14. Auto mode routes each tool call through a classifier that blocks irreversible or destructive commands, rather than relying on user approval prompts. In a controlled study with 1,053 paid testers, humans caught only 13.6% of dangerous commands while auto mode caught 89%. Teams using auto mode ship roughly 25% more pull requests. Enterprise and API users remain opt-in for now.

OpenAI has determined that its upcoming Astra model cannot rule out reaching the Critical cybersecurity threshold under its Preparedness Framework, meaning the model may be capable of identifying and developing functional zero-day exploits in hardened real-world systems without human intervention. In response, OpenAI is implementing isolated testing environments, enhanced model weight protections, universal monitoring for risky actions, and pausing internal Astra activities that don't meet new security controls. The company will also work with government agencies and AI safety organizations to test the model's capabilities.

Anthropic has updated the safety classifiers for Claude Fable 5, reducing biology-related fallbacks by approximately 85%. Previously, Fable 5 routed biology queries to the less capable Opus 5 model due to broad safeguards launched alongside the model. The refined classifier now better distinguishes benign requests—like interpreting lab results or learning about symptoms—from dual-use risks. Dual-use topics including virology, toxicology, and molecular design remain blocked. Anthropic plans trusted access pathways for professional biology researchers and drug developers.

Alibaba has launched Wan3.0 in public beta, a video generation model that doubles the maximum clip length of its predecessor to 30 seconds. Unlike most AI video tools capped at 15 seconds, Wan3.0 accepts multimodal inputs including text, images, video, audio, PDFs, and PowerPoint files, enabling conversion of static documents into dynamic video. The model features high-precision visual continuity, synchronized facial micro-expressions, and an intelligent duration recommendation tool. Users can access it via Alibaba Cloud's Model Studio and Qwen Cloud platforms.

Perplexity has launched Computer for Builders, a platform targeting solo founders and small engineering teams that orchestrates 15+ frontier AI models to automate the full software development lifecycle. Built on its existing Perplexity Computer product, it integrates with GitHub, Datadog, Stripe, Supabase, and Slack to write and deploy code, monitor production, track payments, flag fraud, and send automated weekly growth reports to Slack every Monday. It is available to Pro and Max subscribers.

OpenAI and former Apple designer Jony Ive are developing a hockey puck-sized, doughnut-shaped AI smart speaker expected to launch in 2027 for over $300, according to Bloomberg's Mark Gurman. The battery-powered device has no display and features moving parts that animate when responding, plus lights, a camera, and sensors. Designed to be carried around the home one-handed, it will run ChatGPT voice mode with more advanced models for humanlike interaction and may be the first in a planned family of OpenAI devices.

The accessible source details point to this update: This A.I. Just Created Viruses Not Found in Nature. Because the full article could not be reliably extracted or rewritten, this TLDR stays conservative and is based on headline-level information plus limited source context from nytimes.com.

OpenAI's new country-by-country Signals dataset reveals ChatGPT is shifting from question-answering to task completion, with users at work more than twice as likely to use it for producing outputs like coding, editing, or analysis. Multimedia is the fastest-growing use case globally at 7.8% of messages, exceeding 10% in Brazil and Colombia following the April 2026 ChatGPT Images 2.0 release. Adoption gaps are narrowing, with Peru, Uruguay, and Costa Rica rising fastest in per-capita rankings, and users over 35 now account for a 5% higher share of messages year over year.

OpenAI has updated GPT-5.6 Sol in ChatGPT for Plus and Pro users, delivering more focused answers, fewer factual errors, and a new slider to control how much reasoning the model applies per response. Internally, factual errors dropped 68% compared to GPT-5.5 Instant. Free users get GPT-5.6 Luna as their new default model with unlimited text chats, plus a Think button for harder questions requiring deeper reasoning. Limits on file uploads and images still apply.