Google launches Emergency Location Service (ELS) on Android in Uttar Pradesh, automatically sharing precise user locations with emergency responders during 112 calls. Built into Android 6.0+, ELS uses GPS, Wi-Fi, and cellular data via the Fused Location Framework, achieving ~50-meter accuracy without third-party apps or user setup. Integrated with UP112, it enables real-time location sharing even if calls disconnect, reducing reliance on verbal descriptions. During a pilot, ELS supported over 20

MiniMax launches M2.1, enhancing coding, multilingual, and agent capabilities; excels in VIBE benchmark with 88.6 avg score, supports web, Android, iOS development, and outperforms Claude Sonnet 4.5.

Instacart ends AI-driven dynamic pricing tests after scrutiny over price disparities and fairness concerns; stops use of Eversight tech, commits to transparency amid FTC investigation and $60M refund order.

OpenAI acknowledges prompt injection as an enduring threat to AI agents like Atlas, employing LLM-driven red-team agents and rapid testing to proactively uncover vulnerabilities. Despite layered defenses and user confirmation prompts, experts caution that high-access agents pose significant risks, questioning their current risk-value balance.

GPT-5 reportedly solved an unsolved math problem autonomously, using novel algebraic geometry techniques. The final paper, co-authored with AI models and labeled for transparency, raises questions about attribution, collaboration, and the nature of AI contributions in science.

Sigma Browser launched Eclipse, a privacy-focused browser with an on-device LLM for offline AI. Unlike cloud-dependent rivals, it keeps all data local, ensures uncensored responses, and supports on-device PDF analysis, aiming to make private AI browsing accessible without high technical barriers.

Anthropic launched Bloom, an open-source framework to evaluate AI alignment by testing behaviors like deception and bias across 16 leading models, complementing its Petri tool for scalable, targeted safety assessments.

ChatGPT launches "Your Year With ChatGPT," a personalized annual recap for free, Plus, and Pro users in select English-speaking markets. Featuring badges, a custom poem, and AI-generated art based on user interactions, the privacy-focused feature is available via app or by prompt, with full user control over access.

Kling 2.6, Kuaishou's latest AI model, enables synchronized audio-video generation with advanced voice and motion control. It supports custom voice training for consistent character voices, realistic lip-syncing, and precise full-body motion capture, including complex gestures and facial expressions. The model works with text or image-text inputs, powering diverse applications like vlogs, news, music, and short dramas. Available via third-party platforms at competitive rates (from $0.07/sec), it

Alibaba's Qwen unveils Qwen-Image-Layered, an AI model that splits images into editable RGBA layers for precise element manipulation. Supports resizing, recoloring, object removal, and background swaps. Code and demos available on GitHub, Hugging Face, and ModelScope.

Nvidia unveils NitroGen, an open visual-action AI trained on 40,000 hours of gameplay. Built on GR00T N1.5, it excels across diverse games and virtual environments, achieving 52% higher success in new games. Dataset, code, and model are publicly released.

Google open-sourced Metrax, a JAX library offering standardized, high-performance ML evaluation metrics for classification, NLP, vision, and audio. It supports distributed computing, includes metrics like F1, BLEU, IoU, and SSIM, and uses JAX optimizations (vmap, jit) for efficient "at K" evaluations. A key feature enables computing multiple K-values in one pass, improving efficiency for ranking systems. The release addresses fragmentation in metric implementations across JAX projects and promot

OpenAI lets ChatGPT users customize tone, emoji use, and formatting in new personalization settings, enhancing control over the chatbot's warmth and style.

AWS launched ECS Fargate Spot Mode for simplified, one-step deployment of containerized apps with auto-provisioned ALB, scaling, and TLS. No extra cost; full resource control. Supports HTTPS services, but lacks EC2 launch types and blue/green deployments. Available in all ECS regions.

Cursor acquires Graphite, an AI-powered code review tool detecting vulnerabilities and inefficiencies, to enhance development workflows. Graphite will remain standalone, integrating deeper with Cursor’s ecosystem while advancing AI automation in code reviews.

OpenAI launched GPT-5.2-Codex, its most advanced AI coding model, featuring context compression, enhanced visual understanding, and improved agent reliability. It excels in complex software tasks, benchmarks, and cybersecurity applications, now available to paid users with cautious deployment via a Trusted Access Pilot program.

Google launched Gemini 3 Flash, a cost-efficient model offering Pro-level performance for enterprises. Priced significantly lower than Gemini 3 Pro, it balances speed, accuracy, and cost, supporting multimodal tasks and tool integration. Designed for routine workloads, it reflects an industry trend toward economical AI models while maintaining quality.

Luma AI launched Ray3 Modify, a model that alters videos using character reference images while preserving original performance, expressions, and motion. It enables seamless transitions via start/end frame interpolation and integrates real actors into virtual scenes. Available on Dream Machine, it competes with Runway and Kling.

OpenAI and Anthropic are enhancing youth safety: OpenAI updates ChatGPT with new safeguards for teens, prioritizing safety over intellectual freedom, while Anthropic blocks under-18s from using Claude and reduces AI sycophancy. Both aim to protect minors amid growing regulatory pressure.

Google expands Gemini to verify AI-generated videos using SynthID watermarking, detecting and timestamping AI content in videos up to 100 MB and 90 seconds, available globally.