OpenAI’s AI Cracks a 90-Year Millennium Prize Problem — With Controversy OpenAI Blog & New York Times

An internal OpenAI system produced a proof resolving the Navier–Stokes existence and smoothness problem — one of the seven Millennium Prize Problems and a cornerstone of fluid dynamics — in just 88 hours, delivering both an analytical proof and a Lean formalization. The model identified conditions under which the governing equations completely break down, implying the laws of physics could theoretically fail under extreme states. The announcement was immediately shadowed by controversy: mathematicians noted close parallels to unpublished work by researchers Tristan Buckmaster and Alpöge, and OpenAI denied its agents had accessed that material.

🛡️ Security & Safety

Anthropic Researcher Quits, Warns AI ‘Could Kill All Humans’ — The Verge

Senior safety researcher Jacob Coxon resigned from Anthropic and publicly stated there is more than a 10% chance AI “could kill all humans” by the end of the decade, calling the industry’s race to build self-improving systems a gamble with human existence. Coxon called for binding pacing agreements between labs. The resignation lands the same day Anthropic warned users that hackers are actively stealing Claude subscription tokens to run unauthorized inference.

WeWorm: Zero-Click AI Worm Silently Compromises WeChat Accounts — New York Times

A newly disclosed zero-click exploit called WeWorm can fully compromise WeChat accounts — accessing messages, calls, and linked services — without any action from the target. The attack represents a significant escalation in AI-assisted offensive capability and raises urgent concerns for the platform’s billion-plus users globally.

NSA, CISA & FBI Allege Six Chinese AI Labs Are Running Aggressive Model Distillation Attacks — Via AI Weekly (Advisory AA26-251A)

A joint government advisory alleges DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI have conducted targeted, malicious distillation attacks against US frontier models since late 2024 — systematically replicating capabilities without the underlying compute investment. The advisory marks the sharpest government framing yet of AI capability theft as a national security threat.

Stealing AI Reasoning Traces — Schneier on Security — Researchers demonstrate extracting encrypted reasoning traces from a flagship model by injecting them into a weaker, less-safeguarded sibling model — no jailbreak of the primary model required.

Hackers Are Stealing Claude Tokens from Subscribers — TechCrunch — Anthropic flagged account token theft after a subscriber noticed runaway inference consumption on a dormant account.

I Asked 100 Agents to Hack Me — sshh.io — An experiment using ~100 self-hosted abliterated open-source agents: five accounts compromised, 16 social engineering attempts in five hours — a stress test of near-future autonomous attack scale.

🔬 Frontier Models & Research

Google DeepMind AlphaGenome Atlas Maps All 9 Billion DNA Variants — Google DeepMind

Google DeepMind released AlphaGenome Atlas, a 1-petabyte database predicting the regulatory impact of every possible single-nucleotide variant in the human genome. Researchers can query any variant via a single-impact score without writing code or running the model — potentially compressing years of disease research into interactive lookups.

Mercury 2.5: Largest Diffusion Language Model, at 1,107 Tokens/sec — Inception Labs

Mercury 2.5 — the biggest diffusion LM trained to date — matches cost-optimized frontier models including GPT-5.6 Luna, Gemini 3.5 Flash-Lite, and Claude Haiku 4.5, while outputting over 1,100 tokens per second on standard Nvidia GPUs. Its 260K-token context window and aggressive launch pricing ($0.04/$0.15 per million input/output) make it one of the more compelling cost-performance options at release.

Magic Claims >10× Compute-Efficient Pretraining — Magic — The AI coding startup says its recipe is more than 10× more compute-efficient than leading open-weight base models, achieved through algorithmic improvements rather than raw scale — a meaningful data point for labs without frontier-tier compute budgets.

Anthropic Launches Claude Fable 5.1 — Last Week in AI — Fable 5.1 shipped this week; full technical details are sparse, but the model is noted alongside Astra and other releases as a significant update.

Pretraining Progress Is Mostly Coming from Data — Dwarkesh Patel — Analysis of 2019–2025 scaling finds 3.24× more efficiency gains came from data improvements than model architecture changes, with data quality gains mattering more for smaller models.

🤖 AI Agents

Meta Officially Launches Muse Personal AI Agent — Meta

Muse, Meta’s personal AI agent powered by the new Muse Spark model, is now publicly available on iOS, Android, and muse.ai in the US. It can book travel, send email, make purchases, and take real-world actions on behalf of users through a sandboxed Secure VM with a Sentinel oversight agent. Third-party integrations include Gmail, Spotify, OpenTable, Ticketmaster, and Shopify; pricing is free with limits, or $20/$100/month for more capacity.

GitHub HydraFusion Hits Claude Opus 5 Quality at 36–67% Lower Cost — GitHub

GitHub’s runtime multi-model orchestration system dynamically routes between single-model, cascade, and critique workflows across providers. Offline benchmarks show HydraFusion matched or exceeded Claude Opus 5 quality while slashing estimated costs by up to two-thirds — a compelling case for adaptive model routing in production coding agents.

Instinct AI Assistant Gets Its Own Email Address — TechCrunch — Instinct can now autonomously create and manage email accounts, contact businesses, and handle support requests on users’ behalf.

Instacart Launches Clementine AI Grocery Assistant — TechCrunch

Shipt Adds Conversational AI Shopping Assistant — TechCrunch

📱 Hardware & Consumer

Apple Unveils Foldable iPhone Duo at Fall Event — TechCrunch

Apple’s first foldable phone — the iPhone Duo — arrived today at $2,199, featuring a 7.8-inch interior display and a hinge developed using AI and 3D printing. The lineup also includes the iPhone 18 Pro, which introduces “Reference Image,” a feature that cryptographically signs every captured pixel to prove photos are unmanipulated by AI, and an always-listening Apple Watch. Apple CEO John Ternus positioned the iPhone as the premier AI device, citing on-device model privacy advantages.

Cognition Raises $2B at $48B Valuation — TechCrunch

Andreessen Horowitz, Accel, Founders Fund, General Catalyst, and Avenir backed Cognition at a valuation higher than Cursor’s pre-SpaceX acquisition multiple — a signal that VCs still see room for multiple major winners in AI coding. The company is tracking toward $4–5B in annualized revenue by year-end, though total burn could reach $800M in 2026.

Suno v6 Trained Exclusively on Licensed Music Amid Copyright Suits — TechCrunch — Facing multiple lawsuits, Suno replaced all prior models with Suno v6, trained only on music licensed for AI use.

Amazon Prime Video Launches AI Lip-Sync Dubbing — The Verge

🏢 Enterprise & Policy

Microsoft Agrees to AI Privacy Rules for Schools with AFT — The Verge

Microsoft reached a binding agreement with the American Federation of Teachers and the United Federation of Teachers establishing safety and privacy principles for student-facing AI — arriving one week after two major school systems announced bans on student AI products. The agreement signals growing institutional pressure for enforceable AI guardrails in education.

OECD Study: Students Who Use AI Generally Score Worse — The Verge — Global PISA-linked data shows AI-assisted students underperform non-users overall, though students taught to critically evaluate AI output show modest improvement.

AI Spend Per Employee Slumped at Top Firms in August — TechCrunch — Falling token costs and cheaper models drove down per-employee AI spend at leading enterprises, raising questions about whether adoption is plateauing or just becoming more cost-efficient.

Sequoia Doubles Down on Cymphony for AI Agent Identity Security — TechCrunch — Cymphony gives security teams a unified view of human employees, AI agents, and nonhuman identities alongside their data access.

Generated by Claude Sonnet 4.6 on 2026-09-09T10:00:00Z