AI Frontier · Aug 20, 2026
This edition covers items published over the past 7 days by three sources: OpenAI, Google DeepMind, and Google AI. The three AI security sources — HiddenLayer Research, Lakera Blog, and Protect AI Blog — again had no fetchable content this week (no RSS; a known gap, not an oversight in this edition).
Last 48 Hours
OpenAI: Replit launches Free Mode, powered by GPT-5.6 Luna (Aug 19): the new Free Mode lets users turn ideas into working software without worrying about token costs. Link
OpenAI: strengthening democratic oversight in national security (Aug 18): OpenAI launches a new program providing government agencies with tools, training, and expert support to strengthen democratic oversight of AI use in national security settings. Link
OpenAI: "Pacing model development in an era of cyber-critical capabilities" (Aug 18): OpenAI details how it is strengthening monitoring, alignment, and safety mechanisms for frontier models and adjusting its model release cadence accordingly — the most directly security-relevant piece this week. Link
Aug 17 — Aug 13
OpenAI: "The Defender's Window" (Aug 17): examines how AI is simultaneously reshaping cybersecurity on both offense and defense, covering OpenAI's own security hardening measures and the actions security teams can take today. Link
OpenAI: funding 14 "intelligence age" policy research projects (Aug 17): grants for independent policy research on expanding economic opportunity and strengthening societal resilience. Link
Google DeepMind: releases Gemini 3.7 Flash (Aug 13): the DeepMind blog formally introduces the next-generation lightweight model Gemini 3.7 Flash (the original carried no detailed write-up). Link
OpenAI: "The builder's guide to GPT-5.6" (Aug 13): covers how startups can build AI agents faster and more economically using GPT-5.6 together with smarter model selection and the new Responses API. Link
OpenAI: previews Ultrafast mode, raising GPT-5.6 Sol inference speed by up to 14x (Aug 13): a new API service tier built on Cerebras compute, with output speeds reaching up to 750 tokens per second. Link
This edition excludes purely marketing or commercial-partnership content (such as the ChatGPT Ads expansion into Europe, the Google Gemini and Pixel football club partnerships, and the NVIDIA/Asana customer case studies), focusing instead on model capability and security-related developments.