The day's tech, sifted: Sep 9, 2026

What matters today: Meta launched Muse, a personal AI agent that can book travel and negotiate bills through your own accounts, gated by a Sentinel approval layer and backed by a bug bounty paying up to $300,000 for prompt injection attacks. A credit dispute deepened around OpenAI's Navier-Stokes proof: mathematician Tristan Buckmaster says OpenAI fought dirty after word of his unpublished work reached the company, and Terence Tao weighed in on open problems being "non-renewably mined" by AI. And inside Anthropic, pretraining researcher Jacob Coxon quit the AI industry entirely, warning labs are "gambling with our lives," prompting alignment lead Evan Hubinger to publicly agree he sees a >10% chance AI kills everyone within a decade.

AI / LLMs

• Meta launched Muse, a personal AI agent on web, iOS, Android and WhatsApp powered by Muse Spark 1.3, that can send emails, book tickets and shop with approval; sensitive actions route through a Sentinel dialog that swaps in real credentials only at the network boundary, and Meta opened its bug bounty to everyone with rewards up to $300,000.
• OpenAI shipped ChatGPT Images 2.5, cutting generation latency up to 50% and adding Sketch, a draw-inside-the-chat reference tool invoked with "@Sketch," alongside API models GPT-Image-2.5 Flare and Sunburst.
• The Navier-Stokes credit dispute escalated: NYU's Tristan Buckmaster alleges OpenAI's model finished his and Anthropic researcher Levent Alpöge's unpublished proof days after word of their method reached the company; OpenAI denies it, and Terence Tao publicly praised Buckmaster and Alpöge's work while warning that open problems are increasingly "mined" by AI before their authors can publish.
• 1Password says Codex cut its median pull request cycle time by 11% and lifted engineering productivity almost 21% across its developer teams, per OpenAI's own case study.
• World Labs unveiled Atlas, an omni world model unifying text, image, video and 3D reconstruction into one 3D-grounded spatial context.

Devtools & Infra

• Samsung and TSMC committed to ASML's $400M High-NA EUV machines, becoming its second and third customers after Intel, and joined an industry push to move photomasks from 6-inch to 12-inch, targeting a 40% throughput gain by 2033.
• Argonaut Labs got the full 2.8-trillion-parameter Kimi K3 running at 1 token/s on a single MacBook Pro, streaming all 16 unpruned experts on demand from four SSDs with router-aware prefetch to hide read latency.

Security & Privacy

• Microsoft's September Patch Tuesday fixed a record 972 vulnerabilities, 112 of them critical, including two zero-days already under active exploitation in the Windows Update Stack and Windows ALPC.
• NSA, CISA and FBI issued a joint advisory naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI as running industrial-scale knowledge-distillation campaigns against Claude, ChatGPT, Gemini and Grok since late 2024 to shortcut their own model training.
• Anthropic pretraining researcher Jacob Coxon resigned and quit the AI industry entirely, saying labs are "racing straight to self-improving superintelligence and gambling with our lives"; alignment lead Evan Hubinger replied that he agrees, estimating >10% odds AI kills all humans within a decade and noting Anthropic has no plan yet to solve alignment for superintelligence.
• Police departments across the US are circulating internal memos warning that Meta's camera glasses could be used to secretly record officers, with NYPD counterterrorism and DHS fusion centers flagging the risk to cell layouts and patrol routines; ICE has barred the glasses from federal workspaces as unauthorized body cameras.

Startups & Industry

• A Claude subscriber class action against Anthropic expanded, alleging the Max 5x and Max 20x plans deliver roughly 3.5x and 6-8x Pro-tier usage rather than the advertised 5x and 20x.
• Anthropic quit the Information Technology Industry Council after the trade group lobbied to strip three chip export-control measures, including the AI OVERWATCH and Chip Security Acts, from the defense policy bill.
• Chime agreed to acquire its longtime banking partner Stride Bank for $590M, folding it in as a wholly owned subsidiary; shares jumped over 8% after hours.
• More than 10 US states have rolled back tech tax breaks worth over $1B a year in some cases, part of a wider legislative backlash against data center buildouts.

Research

• A Harvard team predicted 75% of suicide attempts and 87% of suicide-related events in the week before they happened, using six-times-daily smartphone surveys with 600+ high-risk adults and adolescents over three months, far ahead of prior models built around six-month to multi-year horizons.

Elsewhere

• A NATS technical fault grounded UK air traffic for nearly four hours, canceling over 1,000 flights including at Heathrow and Gatwick and stranding some passengers on taxiways for five-plus hours.

Hacker News

Building materials and hobby engineering share the spotlight: How to build a printer walks through building an inkjet from raw parts, and The Helicopter with Radioactive Blades covers a Cold War-era design that used radioisotope tips for lift, a genuinely strange bit of aerospace history.

On the AI/ML side, Mercury 2.5 is Inception Labs' latest diffusion-based language model, pitched on speed rather than benchmark chasing; a Show HN, LLM Attention Visualization, lets you inspect attention patterns directly in the browser and drew a solid front-page audience for a tool post. A paper on LLMs developing novel social biases through adaptive exploration argues models can pick up biased behavior patterns purely from exploration dynamics, not just training data, worth a skeptical read given it's a submission link rather than a published result.

The day's other big boards, Muse, ChatGPT Images 2.5, the Tao/OpenAI Navier-Stokes credit dispute, Kimi K3, the Anthropic resignation thread and Meta smart glasses, are covered above; the South Park "renaming" tweet is a promo bit from the show's own account, not real news.

Threads

• AI safety alarm keeps escalating from inside the labs that build it: Jacob Coxon quit Anthropic warning of a race to superintelligence, alignment lead Evan Hubinger publicly agreed, and both echo yesterday's thread of Pachocki's "An Alien Mind" and Astra's own system card admitting weaker chain-of-thought monitoring.
• Credit and authorship disputes are catching up with AI's output: Buckmaster accuses OpenAI of finishing his unpublished Navier-Stokes work days after word reached the company, the same week Terence Tao warns open math problems are being mined faster than researchers can publish them.
• Consumer AI's agent race picked a battleground: Meta's Muse and OpenAI's ChatGPT Images 2.5 landed the same week, one betting on autonomous action with your accounts, the other on faster, more precise image generation.
• The US and China are racing the AI compute gap from opposite ends: Samsung and TSMC locked in $400M-a-machine ASML tools years ahead of deployment, while federal agencies accuse Chinese labs of distilling US frontier models to shortcut exactly that kind of capital spend.