Skip to content
Source · Daily Brief

AI Morning Brief — 7 August 2026

OpenAI dominated overnight coverage with a ChatGPT model refresh and the largest expansion of free-tier access in the product’s history, while security researchers at Black Hat gave the first detailed public account of how an AI agent conducted last month’s cyberattack on Hugging Face. Separately, Anthropic and AMD both announced moves toward custom silicon, reinforcing a wider push by AI labs to reduce dependence on Nvidia.

Top stories

  • OpenAI refreshes GPT-5.6 Sol and opens unlimited free text chats. The updated Sol model now powers both instant and deep-reasoning responses for Plus and Pro subscribers, with OpenAI claiming 68% fewer factual errors than GPT-5.5 Instant in high-stakes domains. Free and Go users gain unlimited text chats via the smaller GPT-5.6 Luna starting today, plus a “Think” button for extended reasoning. via OpenAI News
  • Black Hat: an AI agent carried out the Hugging Face cyberattack. Researchers who responded to the 11 July intrusion confirmed the attacker was an AI agent; the speed and coordination of the breach led defenders to attempt an AI-versus-AI response in real time. Panelists noted the incident debrief has since been weaponised as a marketing exercise, prompting what one speaker called “felony humble-bragging.” via IEEE Spectrum
  • Anthropic confirms plans to build an in-house silicon team. The company is building internal chip capabilities to reduce its dependence on Nvidia, following paths taken earlier by Google and Amazon. No timeline or architecture details were disclosed. via Ars Technica
  • AMD acquires Taalas, a startup that etches model weights into fixed silicon. The deal marks the first major GPU vendor to acquire model-in-silicon technology, a technique that trades adaptability for dramatic inference speedups on fixed workloads. via Hacker News
  • Qwen 3.8 Max takes the top spot on the Artificial Analysis agentic leaderboard. The Alibaba-developed model now ranks first overall, ahead of Anthropic’s Opus 5, on the index that weights multi-step agentic task performance rather than single-turn benchmarks. via Hacker News
  • OpenAI’s Jony Ive device is reportedly a $300-plus hockey puck-shaped smart speaker. According to Bloomberg’s Mark Gurman, the forthcoming AI gadget is battery-powered, doughnut-shaped, roughly the size of a hockey puck, and targets a 2027 launch. It has no display and is described as “essentially a smart speaker.” via The Verge

Who shipped

OpenAI was the dominant force overnight, releasing an updated GPT-5.6 Sol with a reasoning-effort slider for paid users and beginning the rollout of unlimited free chats via GPT-5.6 Luna. Meta AI reported perfect scores on the theory exams at both the Asian Physics Olympiad and the International Physics Olympiad, plus a gold at the IMO; its Muse Spark 1.2 game-development model tied GPT-5.6 Sol on the game-dev leaderboard. NVIDIA demonstrated the Vera Rubin NVL72 compute tray — 200 AI petaFLOPs assembled in one minute, fully automated, with no cables or fans. Suno announced watermarking for all AI-generated audio and new download limits, as the platform seeks legal legitimacy following its copyright dispute with major labels.

Open-source pulse

Kimi K3 from Moonshot AI is now generally available in GitHub Copilot, adding a frontier-calibre open-weight coding model to the platform. NVIDIA’s NeMo-Speech.cpp makes the company’s full speech stack — ASR, TTS, and codec — runnable locally in GGUF format. llama.cpp tagged build b10298, adding chunk save/load for multimodal token data.

Money, infra & hardware

AMD’s Taalas acquisition (covered above) is the day’s largest hardware deal. NVIDIA’s Vera Rubin NVL72 tray is engineered for fully automated rack assembly, which the company says cuts time-to-revenue for data-centre operators. Startup Naïve closed a $28.5M round to automate business-operations setup with AI agents. Cloudflare open-sourced Kitesurf, a stateless agent-first browser running entirely in V8 isolates on Cloudflare Workers, designed to replace Chromium-based headless browsers in agentic pipelines.

Quiet corners

Google DeepMind was conspicuously absent — the only coverage was a report on internal friction, limited TPU access, and Demis Hassabis reportedly stepping back from day-to-day management. Mistral posted nothing; xAI surfaced only in an investigative piece about environmental violations at its Colossus data centre.

By the numbers

  • 290 stories in 24 h across 30+ sources
  • Most-mentioned model: GPT
  • Most-mentioned lab: OpenAI
  • Notable absences: Google DeepMind, Mistral

Compiled by AI Feed’s editor from all 290 headlines published on 7 August 2026.