AI News Flash for September 08, 2026 at 21:00 UTC.
Topics:
1. Meta launches Muse personal AI agent in the US = free + $20 Power / $100 Maximum, iOS/Android/muse.ai/WhatsApp, 18+, Muse Spark 1.3, Secure VM + Sentinel, Link purchase protection, Confidential VM later 2026, Zuckerberg/Wang/Shah quotes
https://techcrunch.com/2026/09/08/meta-debuts-its-muse-ai-agent-will-consumers-trust-it/
2. NSA/CISA/FBI joint advisory AA26-251A (Sept 8) = six China-based AI firms (DeepSeek/Moonshot/Alibaba/MiniMax/StepFun/Z.AI) industrial-scale distillation, billions of tokens since late 2024, likely with Chinese government awareness, transfer stations / tens of thousands of fraudulent accounts / shared premium subscriptions, DeepSeek $5.6M misleading, Moonshot Kimi-K3 from Claude Fable 5, Z.AI from GPT-5.5/Opus 4.8, three mitigations, Andersen quote, no sanctions
https://www.cisa.gov/news-events/cybersecurity-advisories/aa26-251a
3. Google Cloud CEO Kurian at Goldman Sachs Communacopia (Sept 8) = TPU business more than twice next-largest hyperscaler, three silicon types, price-performance 2.7x training / 80% inference / 30% CPU, Q2 revenue $24.77B +82%, backlog $514B (Q1 $460B), 2026 capex $195-205B, Q2 capex $44.9B +100%, Morgan Stanley Nowak TPU $84B in 2027 (0.3/3.2/4.2 GW), stock flat ~$338
https://www.investors.com/news/technology/google-stock-cloud-kurian-ai-chip-business/
4. OpenAI ChatGPT Images 2.5 (Sept 8) = Sketch (@Sketch), templates, image comments, prompt sharing, latency up to -50%, API GPT-Image-2.5 Flare/Sunburst, pricing image input $8 / output $30 / text input $5 per 1M tokens, unsafe rate 1.09%/1.41% vs 1.64%, all ChatGPT/Work/Codex users
https://openai.com/index/introducing-chatgpt-images-2-5/
5. NVIDIA CUDA Rust (Sept 8 Technical Blog) = cuda-oxide (SIMT, rustc backend MIR to Pliron to LLVM to PTX, early alpha, nightly-2026-04-03, CC 8.0+/CUDA 12.x) + cutile-rs (Tile, CUDA Tile IR JIT, crates.io, stable Rust 1.89+/CUDA 13.3, used by HF Grout / mistral.rs), borrow checker rejects aliasing, C++/Python interop planned, RustConf 2026 Montreal Sept 8-11 Elibol talk
https://developer.nvidia.com/blog/introducing-cuda-rust-two-tracks-for-writing-gpu-kernels